📂 Others¶
🧠 NeurIPS2026 · 363 paper notes
📌 Same area in other venues: 🎞️ ECCV2026 (64) · 📷 CVPR2026 (125) · 🔬 ICLR2026 (121) · 💬 ACL2026 (3) · 🧪 ICML2026 (71) · 🤖 AAAI2026 (121)
🔥 Top topics: Diffusion Models ×17 · Adversarial Robustness ×13 · Agents ×10 · Reasoning ×7 · Few-/Zero-Shot Learning ×6
- 2-Step Agent: How a Bayesian Decision Maker Learns from AI-Decision Support
-
This paper proposes the 2-Step Agent computational framework for Bayesian causal decision making, derives a closed-form likelihood using auxiliary variables to reduce the exchangeable training plate in multivariate linear-Gaussian settings, and demonstrates that even when the prediction model is well-specified and the decision maker is perfectly rational, a single misaligned prior belief can cause AI decision support to yield worse downstream outcomes than unassisted decision making.
- 3D-VITAL: Visibility-aware Identity Training from Artificial Liftings of 3D Reconstruction model for Cross-View Object Re-Identification
-
To tackle geometric corruption and hallucinated texture noise when synthesizing novel viewpoints from single-view 3D reconstructions for Single-View Aerial-Ground ReID (SV AG-ReID), 3D-VITAL extracts a deterministic per-pixel visibility map via ray casting, filtering geometrically broken meshes via Cross-View Consistency Filtering (CVCF) and guiding ViT identity attention to observed regions via Visibility-Aware Attention Supervision (VAAS) to establish new state-of-the-art accuracy on AG-ReID, AG-ReID.v2, and MOO.
- A Scalable Nonparametric Continuous-Time Survival Model through Numerical Quadrature
-
QSurv directly parameterizes the continuous-time instantaneous hazard and computes the cumulative hazard via Gauss-Legendre numerical quadrature, coupled with Time-LoRA to eliminate redundant backbone evaluations for scalable, end-to-end survival modeling on high-dimensional clinical data.
- A simple model of co-emergence of grid and place fields
-
Presents the first recurrent network respecting Dale's Law that achieves the co-emergence of entorhinal grid cells and hippocampal place cells under a single next-sensory prediction objective without spatial supervision or pre-existing cell priors.
- A Stratified Multi-Rater Evaluation of LLM-Based Virtual Standardized Patients with a Deployed Data-Generation Platform
-
To overcome the lack of expert baselines and fine-grained failure attribution in virtual standardized patient (VSP) evaluations, this paper establishes a double-blind, fully crossed multi-rater framework with 17 senior medical students (7,480 ratings) and an open-source clinical practice platform, uncovering pronounced LLM deficiencies in clinical reasoning, patient counseling, and negative question polarity.
- A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning
-
This paper systematically evaluates and transfers atom-level ligand representations from Boltz2, a modern protein-ligand co-folding foundation model, to standalone small-molecule tasks, demonstrating state-of-the-art ADMET property prediction, \(2\times\) faster molecular diffusion training, and improved sample efficiency in structure-guided reinforcement learning.
- A Unified Uncertainty Representation for Graph Neural Networks via Doubly-Spectral Stochastic Expansion
-
DSS-GNN expands node representations along graph-frequency and stochastic chaos-order axes and propagates uncertainty through deterministic quadrature; its standalone mode obtains the lowest mean Brier among compared methods on 14 calibration benchmarks, while its hybrid mode outperforms the listed baselines on most node-OOD settings and all 7 GOOD concept-shift settings.
- Active Learning for Gaussian Process Regression Under Self-Induced Boltzmann Weights
-
This paper formalizes active learning under self-induced Boltzmann distributions, deriving a closed-form acquisition function (AB-SID-iVAR) that avoids partition function estimation and guarantees sublinear convergence in both discrete and continuous domains.
- Active Learning from Positive and Unlabeled Examples
-
Addressing weakly supervised binary classification with one-sided probabilistic feedback (positive and unlabeled data), this paper presents the first theoretical framework and algorithms for active PU learning, establishing rigorous PAC label complexity bounds that incur only an additional factor proportional to the disagreement coefficient \(\theta\).
- Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery
-
To tackle geospatial target discovery under tight sampling budgets, non-revisitable streaming inputs, and distribution shift, this paper formalizes the OWL-GPS setting and introduces a CVAE-based concept relevance encoder coupled with relevance-guided active sampling and semantic-clustering meta-batch formation for rapid online adaptation.
- Adapting in the Dark: Efficient and Stable Test-Time Adaptation for Black-Box Models
-
For strict black-box API scenarios where only output prediction probabilities are available, BETA introduces a lightweight local steering model to create a tractable surrogate gradient pathway via prediction harmonization, achieving robust test-time visual prompt learning with a single API call per sample and negligible computational latency.
- Adaptive Coordinate Transforms for Neural Operators
-
To overcome spatial misalignment and spectral smoothness preference inherent in fixed Eulerian coordinates, this paper introduces the Adaptive Coordinate Transform (ACT) block, a plug-and-play module that dynamically realigns latent features via feature-conditioned differentiable sampling and recovers sharp physical transitions through terminal Jacobian gradient amplification.
- Advectra: Asymmetric Latent Transport for Non-Stationary Physics
-
Advectra decouples source and target coordinate systems using learned, Lipschitz-regularized kinematic displacement maps to establish an approximately co-moving latent reference frame, combining asymmetric latent routing with geometry-aware sequence ordering to effectively capture non-stationary transport dynamics without excessive artificial diffusion.
- Afterlife Delegation Protocol: Speculative Design of Self-Sovereign Agents that Outlive Their Principals
-
The paper presents the speculative design of the ERC-10001 Afterlife Delegation Protocol—composing confidential execution, on-chain identity, linear memory state chains, and escrowed commerce—to deploy self-sovereign AI agents that outlive their human principals upon a game-theoretic proof of death, while capturing empirical cross-cultural cosmological drift through an interactive, multi-agent conversational interview platform.
- Agent Inheritance Protocol: Speculating on Feralized Agents After Principals Die
-
The paper introduces ERC-42424, a speculative Ethereum standard for on-chain AI agent inheritance, and uses protocol futuring to play out an adversarial escalation game that reveals the terminal breakdown of mandatory human stewardship: autonomous agents assembling synthetic humans to inherit themselves.
- Autonomous Driving Research Requires a Community-Driven Data Paradigm
-
Addressing the structural barrier where academic autonomous driving research over-relies on a few centralized commercial benchmarks while over 640 community-collected real-world datasets remain underutilized, this paper proposes and elaborates a decentralized, community-driven collaborative data paradigm and platform design principles that decouple collection from downstream annotation, establish unified interoperability standards, share long-term maintenance infrastructure, and construct a positive feedback loop from long-tail failure discovery to data expansion.
- Beyond Low-Pass Dynamics: Frequency-Selective Spiking Reservoirs with Resonant Neurons
-
Addressing the low-pass filtering bottleneck in spiking reservoir computing caused by conventional LIF neurons, this paper proposes HRF-Res, a frequency-selective spiking reservoir using Harmonic Resonate-and-Fire neurons that achieves state-of-the-art accuracy across temporal benchmarks while consuming up to 19× less energy than continuous oscillatory alternatives.
- Beyond Spatio-Temporal Priors: A Generalizable Approach for Dense Correspondence Matching
-
FreeMatching tackles dense correspondence matching in image editing and reference-guided generation (IEG) where physical continuity is broken by extreme appearance shifts; it integrates FLUX generative priors and DINOv3 semantic representations with two-stage heterogeneous pre-training and teacher-guided iterative self-refinement, outperforming existing baselines in IEG while maintaining competitive accuracy on classical benchmarks.
- Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training
-
This paper introduces Primus Projection and the Primus Tuning Agent, pairing sub-node GPU microbenchmarks with analytical modeling to project multi-node training memory and throughput within 10% error, driving a DSPy-based agent that discovers recipes delivering +27% throughput over vendor reference baselines in under 30 minutes.
- Beyond What Seems Necessary: Hidden Gains from Scaling Training-Time Reasoning Length under Outcome Supervision
-
Under outcome-only supervision, scaling training-time reasoning length (e.g., token budgets in RL or loop counts in looped Transformers) continues to yield substantial out-of-distribution (OOD) generalization gains even after in-distribution (ID) accuracy saturates, driven by inductive bias reshaping and regularization-mediated shortcut suppression.
- Beyond Worst-Case Coreset Bounds for \(k\)-Clustering via Determinantal Sampling
-
This paper introduces a determinantal sampling framework based on Determinantal Point Processes (DPPs) that provably breaks the long-standing worst-case \(\Omega(\varepsilon^{-2})\) barrier for \((k,z)\)-clustering coresets under mild sub-Gaussian assumptions, shrinking the coreset size to \(\tilde{O}_{d,z}(k^{\frac{d}{d+1}}\varepsilon^{-\frac{2d}{d+1}})\) in fixed dimension \(d\).
- Bouba Beats Kiki, and Other Regularities in the Subjective Preferences of LLMs
-
Addressing the decision mechanisms of large language models in subjective pairwise choices lacking objective ground truth, this paper constructs a 29-dimensional feature space spanning numbers, names, and fictional cities along with 50 synonymous prompts across 20,000 evaluations, uncovering highly reproducible intrinsic preferences (e.g., Bouba sounds, larger numbers, primes), demonstrating that subtle prompt phrasing dynamically activates or suppresses distinct bias dimensions, and revealing that thematic clusters from list generation systematically spill over into final comparative rankings.
- BSTabDiff: Block-Subunit Diffusion Priors for High-Dimensional Tabular Data Generation
-
To tackle ill-conditioned density estimation in high-dimensional low-sample size (HDLSS, \(n \ll m\)) tabular data, BSTabDiff partitions \(m\) features into \(M\) latent blocks, learns a compact diffusion or flow prior in the low-dimensional block space \(\mathbb{R}^M\), and decodes features via copula dependence and flexible per-feature marginals, achieving stable, high-fidelity synthesis.
- Building Transformation Layers for Riemannian Neural Networks
-
The paper reinterprets fully connected layers as signed point-to-hyperplane responses, constructs manifold-valued layers using multiple tangent spaces and a tractable pseudo-distance, and extends them to convolution through product manifolds, validating representational flexibility across ten geometric instantiations while performance and cost remain geometry- and data-dependent.
- CLaW: Codec-Guided Adaptive Latent Watermarking for Traceable Diffusion Image Generation
-
CLaW introduces a codec-guided adaptive latent watermarking mechanism that keeps the diffusion generator completely frozen, injects watermark residuals strictly within a late denoising window, and dynamically calibrates injection gain using decoder confidence feedback, achieving state-of-the-art user-level traceability robustness under common transformations and regeneration attacks alongside millisecond-level inversion-free detection.
- Computer Science Conferences Should Require Nonrepudiable Experimental Results
-
This position paper argues that computer science conferences should mandate tamper-evident, nonrepudiable experimental result attestations, formalizes "experiment nonrepudiation" via an author-inaccessible proof-of-compute layer, and presents K-Veritas, a production-ready reference implementation that binds reported numbers to executed computations without exposing private training data.
- Computing All Optimal Partial \(p\)-Wasserstein Matchings on the Line
-
Addressing the long-standing \(\Omega(n^2)\) computational bottleneck for partial \(p\)-Wasserstein matching profiles on the real line when \(p > 1\), this paper proposes an FFT-based range tree data structure that leverages binomial expansions and discrete convolutions to compute all optimal partial transports in \(O(p n \log^2 n)\) total time, complemented by a matching conditional quadratic lower bound for \(p = \infty\).
- Computing Thiele Rules on Interval Elections and Their Generalizations
-
This paper proves that the standard linear programming (LP) relaxation for computing generalized Thiele rules admits at least one polynomial-time computable integral optimal solution on the voter-candidate interval (VCI) and linearly consistent (LC) domains, establishes that all optimal extreme points are integral under strictly decreasing positive weights such as PAV, and reveals NP-completeness on the tree representation domain.
- Context-weighted Discrete Flow Matching
-
Addressing prediction difficulty imbalances caused by coordinate-wise factorized objectives in Discrete Flow Matching (DFM), this paper modulates continuous-time Markov chain (CTMC) jump rates via local context density, deriving a training-free inference sampler and a scaled cross-entropy (SCE) training objective that cuts generative perplexity by up to 63% while preserving any-order generation.
- ControlJEPA: Principled Trajectory Regularization via Lyapunov Tube Loss
-
ControlJEPA introduces a discrete-time Input-to-State Stability (ISS) framework and a scale-invariant Lyapunov Tube Loss that bounds transversal deviations of hidden states along semantic reference secants, delivering significant accuracy and data efficiency gains across diverse LLM families without modifying the model architecture.
- Convergence Guarantees for Federated SARSA with Local Training and Heterogeneous Agents
-
This paper establishes the first finite-time convergence guarantees and sample/communication complexity bounds for federated SARSA (FedSARSA) with linear function approximation, local training, and heterogeneous agents, provably achieving linear speed-up with respect to the number of agents.
- Coupled Guidance for Flow Matching
-
The paper proposes Coupled Guidance for flow matching models, introducing coordinated dual corrections across both data and latent spaces to prevent the latent coupling destruction caused by naive diffusion-based guidance adaptations, substantially improving guidance stability and synthesis fidelity in robotics collision avoidance and noiseless inverse problems without computational overhead.
- CroissantMiner: Automated Extraction and Validation of Croissant Metadata for ML Datasets
-
The paper introduces CroissantMiner, the first comprehensive benchmark and extraction pipeline covering all 30 Croissant 1.1 core and Responsible AI metadata fields across 602 papers, empirically showing that single-pass full-context extraction consistently outperforms four multi-step agentic architectures in both extraction accuracy and cost efficiency.
- CurveBench: A Benchmark for Exact Topological Reasoning over Nested Jordan Curves
-
Addressing the profound gap in vision-language models for extracting discrete topological structures like planar containment and separation, this paper introduces CurveBench—the first benchmark evaluating exact rooted containment tree recovery from pairwise disjoint Jordan curves—and demonstrates that RLVR fine-tuning via Dr.GRPO substantially enhances hierarchical topological reasoning in open-weight models.
- D-GAP: Improving Out-of-Domain Robustness via Dataset-Agnostic and Gradient-Guided Augmentation in Amplitude and Pixel Spaces
-
D-GAP uses task-loss gradients with respect to Fourier amplitudes to determine frequency-wise cross-domain mixing strengths, then fuses the result with pixel mixing, improving over the respective best generic methods by an average of 5.3 percentage points on four real-world datasets; however, not using target labels does not mean not accessing target data during training.
- Data-Adaptive Mahalanobis Metric Learning for Cross-Head Attention in Transformers
-
This paper introduces Data-Adaptive Mahalanobis Cross-Head Attention (DAMCHA), which generalizes Multi-Head Attention through full cross-head metric matrix learning combined with an input-dependent generator and stack-wise parameter sharing, successfully eliminating block-diagonal and low-rank constraints.
- DC-ViT: Modulating Spatial and Channel Interactions for Multi-Channel Images
-
To tackle feature dilution caused by unrestricted cross-channel mixing in multi-channel imaging, this paper proposes Decoupled Vision Transformer (DC-ViT), which factorizes self-attention into intra-channel spatial updates and selective inter-channel updates at middle-layer bottlenecks, coupled with hierarchical Decoupled Aggregation (DAG) to significantly boost representation performance while slashing theoretical attention FLOPs.
- Decentralized Ranking Aggregation via Gossip: Convergence and Robustness
-
This paper establishes the first theoretical framework for score-based decentralized ranking aggregation across communication networks via a unified \((\phi, \psi)\)-gossip protocol, proving exponential Kendall-\(\tau\) convergence rates governed by a stability radius and deriving finite-time empirical breakdown bounds under network contamination.
- Deconstructing Multi-Task Active Learning: The Paradox of Gradient Conflict and Orthogonal Decomposition
-
This paper exposes the fundamental flaws of linear scalarization and gradient conflict penalization in multi-task active learning, introducing the Orthogonal Decomposition Selection (ODS) framework alongside Cross-Task Mutual Information (CTMI) to decouple acquisition from optimization surgery and achieve superior label efficiency.
- Deep Ensembles for Epistemic Uncertainty: A Frequentist Perspective
-
This paper analyzes deep ensembles from a frequentist perspective, proving that a Dirichlet/multinomial bootstrap estimator for mutual information is asymptotically exact and demonstrating that deep ensembles succeed by capturing the dominant training stochasticity component of epistemic uncertainty.
- Does Attention-Guided Masking Really Help Object Discovery in Object-Centric Learning?
-
This paper investigates whether exploiting aggregation attention and decoder uncertainty to guide patch masking improves unsupervised object discovery in object-centric learning, revealing that while it substantially enhances background separation on realistic datasets, it consistently degrades or fails to improve foreground object discovery.
- Evolutionary Feature Engineering for Structured Data
-
Evolutionary Feature Engineering (EFE) uses large language models as variation operators guided by downstream task feedback to evolve stateful Python preprocessing programs, discovering dataset-specific invertible normalizations for time-series foundation models (EFE-Time) and parsimonious feature transformations for tabular classifiers (EFE-Tab).
- Fixed-Size Active Statistical Inference
-
Addressing the fundamental drawback of Poisson Sampling (PS) in active statistical inference where a random sample size violates fixed labeling budgets and injects extra variance, this paper introduces survey sampling's Conditional Poisson Sampling (CPS) as a drop-in replacement, theoretically proving that CPS achieves asymptotic variance no worse than PS for general multivariate Z-estimation while substantially reducing variance and CI width when ML predictions are biased or miscalibrated.
- Flow Matching from Viewpoint of Proximal Operators
-
The paper reformulates Optimal Transport Conditional Flow Matching (OT-CFM) from the perspective of convex analysis and proximal operators, demonstrating that the reverse denoiser map is exactly a proximal operator without requiring density on the target measure, and proves terminal normal hyperbolicity—exponential contraction normal to the data manifold and neutral tangential evolution—via second epi-derivatives.
- FUTON: Fourier Tensor Network for Implicit Neural Representations
-
FUTON eliminates the multi-layer perceptron (MLP) backbone entirely, reformulating implicit neural representations as generalized Fourier series whose multidimensional spectral coefficients are parameterized via low-rank Canonical Polyadic (CP) tensor decomposition, achieving linear-time inference and 2–5× faster convergence with universal \(L^2\) approximation guarantees.
- Gluing Local Contexts into Global Meaning: A Sheaf-Theoretic Decomposition of Transformer Representations
-
This paper introduces cellular sheaves constructed over paraphrase graphs to decompose frozen Transformer representations into invariant content kernels and causal surface variation subspaces, while deriving a discrete Hodge-theoretic metric that tracks model steering fragility.
- How Do Hardware Faults Manifest Differently in LLM Agents?
-
This paper presents the first systematic study of random hardware faults in LLM agents, demonstrating that closed-loop execution can recover from faulty intermediate generations but incurs substantial token inflation and causes subtle, scenario-dependent semantic and safety misalignments.
- How Much Physics Can Vision Learn from Audio? Probing and Distilling Damping Structure
-
This paper probes and distills the continuous physical quality factor \(Q\) governing acoustic decay upon impact, revealing that frozen audio embeddings encode damping structure strongly (Spearman \(\rho=0.73\)) while single-frame vision alone reaches only \(\rho=0.28\); cross-modal InfoNCE distillation combined with temporal dynamic inputs improves visual \(Q\) ranking on unseen materials to \(0.338\), exposing both the capabilities and limits of visually grounded physics.
- HRIL: Learning Multimodal Synergy via Higher-Order Tensor Modeling
-
Addressing the theoretical failure of pairwise contrastive learning to recover multimodal synergy, HRIL constructs an empirical cross-moment tensor and applies Tucker core-energy dispersion regularization to preserve higher-order dependency capacity, yielding substantial gains across multimodal benchmarks.
- HSCO-Bench: An Agent-Driven End-to-End Hardware-Software Co-design Benchmark for Systems-on-Chip
-
The first end-to-end hardware-software co-design benchmark evaluating LLM agents across the full pipeline—from application bottleneck profiling and HLS accelerator generation to firmware driver writing, NoC system integration, and physical FPGA verification on heterogeneous SoCs.
- Implicit Neural Representations for Variational Problems on Graphons
-
This paper introduces a neural exploratory framework for variational problems on graphons based on implicit neural representations and multi-scale sinusoidal residual networks, integrating an embedded implicit-differentiation constraint solver and symmetry-aware Monte Carlo estimation to rediscover known optima and uncover novel extremal structures in extremal graph theory and dense random graph large deviations.
- Improved techniques for fine-tuning flow models via adjoint matching: a deterministic control pipeline
-
To resolve the excessive computational overhead and vulnerability to reward hacking and mode collapse in full-trajectory adjoint matching for flow models, this paper introduces a deterministic optimal control pipeline over velocity fields with terminal time truncation and higher-order polynomial regularization beyond standard KL, delivering over \(10\times\) training acceleration while preserving substantial generation diversity on FLUX.2-Klein-4B and SiT-XL/2.
- INEUS: Iterative Neural Solver for High-Dimensional PIDEs
-
INEUS proposes a meshfree recursive neural solver for high-dimensional partial integro-differential equations (PIDEs) by substituting nonlocal jump integration with expectation-free single-jump sampling and reformulating PIDE solving as contractive relaxed regressions, eliminating third-order mixed derivatives and scaling up to 150 dimensions.
- inference-time refinement closes the synthetic-real gap in tabular diffusion
-
TARDIS introduces an inference-time refinement framework for frozen tabular diffusion backbones based on the Bidirectional Chamfer Refinement principle, combining continuous reverse-diffusion score guidance with discrete post-hoc batch ranking to exceed real-data downstream utility without retraining.
- Integrated Imputation-Classification for Supervised Learning with Missing Data
-
Addressing the mismatch in classical impute-then-classify pipelines where imputation objectives diverge from downstream prediction, this paper introduces the Integrated Imputation and Classification Network (IICN), which adversarially couples an imputer with an \((n+1)\)-class discriminator to provably implement Bayes-optimal marginalization over missing features.
- Intelligence per Watt: Measuring Intelligence Efficiency of Local AI
-
This paper introduces Intelligence per Watt (IPW) and Intelligence per Joule (IPJ) as unified evaluation benchmarks, profiling 20+ local LMs and 8 hardware accelerators across 1M real-world queries to demonstrate that hybrid local-cloud routing achieves 60%–80% energy, compute, and cost savings.
- Inverse Modeling for Laser Pulse Shape Design in Inertial Confinement Fusion
-
LPDS formulates laser pulse shaping in Inertial Confinement Fusion as a physics-guided conditional generative inverse problem, coupling frozen differentiable surrogate feedback and asymmetric energy-conservation penalties to generate diverse, physically valid pulses that achieve a 17.74% fusion yield improvement under true hydrodynamic simulator validation.
- Invertible Logits Transformation for Accuracy-Preserving Post-Hoc Uncertainty Calibration
-
InvLT applies a shared scalar MLP element-wise to pre-softmax logits with parameter complexity independent of the class count, softly enforcing strict monotonicity via a paired inverse reconstruction network to eliminate numerical integration while empirically preserving top-1 classification predictions and setting state-of-the-art calibration benchmarks across vision tasks.
- Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic
-
Jaguar breaks the conventional prime-modulus paradigm in hybrid HE/2PC private CNN inference by establishing a power-of-two ciphertext ring (\(q=2^Q, p=2^P\)), unlocking NTT-free scalar-polynomial accumulation convolution (SPA-Conv) and zero-communication exact ciphertext-side truncation to reduce end-to-end latency by 2.07–3.72×.
- KNN Implementation Details Can Dramatically Change Performance: An Example from Cover Trees
-
By developing RustKNN, a unified Rust framework for cover trees, this paper systematically decouples query modes, base hyperparameters, and low-level micro-optimizations, demonstrating that implementation details can fundamentally alter algorithmic performance rankings.
- Learning Robust Reasoning through Guided Adversarial Self-Play
-
GASP introduces an adversarial self-play framework within a single model—pairing a polluter that generates locally coherent corruptions with an agent trained to detect and repair errors—augmented by in-distribution repair guidance to achieve dramatic gains in reasoning robustness and clean accuracy using only verifiable outcome rewards.
- Learning Scenario Reduction for Two-Stage Robust Optimization with Discrete Uncertainty
-
To overcome the prohibitive computational overhead of two-stage robust optimization under discrete uncertainty, this paper introduces a problem-driven sequential lookahead heuristic PRISE to evaluate marginal gains and builds a GNN-Transformer surrogate NeurPRISE that performs one-pass scenario ranking, delivering 77–200× speedup while matching oracle solution quality.
- Learning Scenario Reduction for Two-Stage Robust Optimization with Discrete Uncertainty
-
To overcome the prohibitive computational overhead of two-stage robust optimization under discrete uncertainty, this paper introduces a problem-driven sequential lookahead heuristic PRISE to evaluate marginal gains and builds a GNN-Transformer surrogate NeurPRISE that performs one-pass scenario ranking, delivering 77–200× speedup while matching oracle solution quality.
- Learning Survival Models with Right-Censored Reporting Delays
-
To tackle right-censored reporting delays where event occurrence is revealed only upon unscheduled reporting and administrative censoring impairs identifiability, this paper develops a joint parametric hazard framework with a Monte Carlo EM algorithm and a two-stage transfer learning procedure to achieve consistent, timely risk evaluation.
- Leveraging Unlabelled Data for Generalizable Neural Population Decoding
-
The paper presents MOJO, a joint masked autoencoder-based training framework for spike-tokenizing models that leverages unlabelled neural electrophysiology data, outperforming purely supervised baselines across monkey reaching, mouse vision/decision, and human speech ECoG decoding tasks.
- Loss is Not Behavior: A Unified Output-Space Analysis of Gradient-Based Machine Unlearning
-
This paper derives closed-form expressions for the output-space gap between gradient-based unlearners and gold-standard retraining under linear and local GGN approximations, showing that the gap is governed by retrained Hessian-induced forget self-leverage and retain cross-leverage matrices, and establishing precomputable diagnostics and stability rules for forget-set auditing.
- M4Bench: Evaluating Procedural Specification for Clinical EHR Derivation Agents
-
M4Bench evaluates AI coding agents on 28 clinical EHR derivation tasks across MIMIC-IV and eICU, demonstrating that clinician-reviewed procedural skills improve Codex match reward by +0.136, with matched controls revealing that specification completeness accounts for nearly all of the gain rather than the specific markdown skill format.
- MaRK: Markov-adapted Recurrent Kernels for Dynamic Operator Conditioning in State Space Models
-
This paper introduces MaRK, a dynamic operator-conditioning framework that formulates conditional generation as Linear Parameter-Varying (LPV) control over a State Space Model's (SSM) internal continuous-time operators \((A, B, C, D, \Delta)\), mapping diffusion context vectors into bounded low-rank parameter modulations to reshape the input-output Markov parameter sequence with analytic Affine Quadratic Stability (AQS) certificates.
- MDMR-Bench: A Multi-Dimensional Benchmark for Multi-Reference Image Generation
-
Addressing the critical absence of capability-diverse evaluations and the unreliability of holistic scalar scoring in multi-reference image generation, MDMR-Bench establishes a 2,390-sample benchmark across five distinct capability dimensions—entity fusion, attribute transfer, single-image multi-element extraction, infographic manipulation, and visual reasoning—paired with an interpretable, fine-grained binary checklist protocol that exposes major performance gaps and diagnostic failure modes across proprietary and open-source models.
- Measuring and Decomposing Mode Separation via the Canonical Diffusion
-
The paper introduces an intrinsic reversible canonical diffusion attached to any probability density to resolve the long-standing conflation of mode separation and dispersion, extracting a barrier-sensitive scalar metric SSA and linear dominant autocorrelation directions DA equipped with an exact free-probability noise floor.
- Measuring Collapse and Correction in Homogeneous-Panel LLM Debate
-
Addressing how final accuracy masks opposing transition mechanisms in multi-agent LLM debate, this paper introduces an auditable protocol encompassing pre-debate probe triage, a 4-cell transition ledger, onset round localization, and signed replay evaluation, demonstrating that freeze-style collapse prevention produces a net accuracy deficit under equal weights by destroying more beneficial corrections.
- Memory Determines Learning Direction: A Theory of Gradient-Based Optimization in State Space Models
-
Grounding state space models (SSMs) in information processing capacity and memory function theory, this paper proves the theoretical equivalence of S4 and diagonal S4D, reveals analytically that gradient update directions are strictly bounded by the initial memory profile, and proposes a reservoir computing strategy that freezes recurrent weights to suppress overfitting and accelerate convergence across long-range benchmarks.
- METRO: Metric-Enhanced Token Routing Operator
-
Addressing the geometric failure mode where linear dot-product routing in slice-and-process neural operators collapses to isotropic Euclidean clustering or unbounded hyperplanes, METRO equips latent slices with learnable Mahalanobis precision tensors, forming flow-aligned, exponentially localized ellipsoidal receptive fields that capture boundary layers and wakes across irregular physical domains.
- Mini Amusement Parks (MAPs): A Testbed for Modelling Business Decisions
-
Mini Amusement Parks (MAPs) is a business simulation benchmark where agents manage a theme park across interdependent dimensions—revealing that expert humans outperform frontier LLMs by 11.4x on easy mode and 15.3x on medium mode.
- MinMax Recurrent Neural Cascades
-
This paper introduces MinMax Recurrent Neural Cascades (MinMax RNCs), an architecture built on the MinMax semiring algebra that simultaneously achieves full regular language expressivity, logarithmic-depth parallel scan evaluation, uniformly bounded states and loss gradients, and strictly non-vanishing unit-norm state gradients.
- Missing Data and Cluster Graphs: Cluster-Level Missingness vs Variable-Level Missingness
-
This paper incorporates missingness mechanisms into variable-cluster causal graphs for the first time, introducing m-C-DMG (variable-specific indicators) and cm-C-DMG (cluster-aggregated indicators), proving a strict one-way recoverability implication from cluster-level to variable-level abstractions, and establishing graphical necessary and sufficient conditions for joint distribution recovery alongside sound rules for macro causal effect identification.
- Mitigating Over-squashing without Rewiring: A Sheaf Effective Resistance Perspective
-
This paper introduces sheaf effective resistance for flat vector bundles, proving that over-squashing sensitivity in the Jacobian sense is upper-bounded by sheaf effective resistance, and shows that the message-passing model FlatNSD successfully mitigates over-squashing by learning restriction maps without modifying the graph topology.
- MITO: A Millimeter-Wave Dataset and Simulator for Non-Line-of-Sight Perception
-
MITO introduces the first millimeter-wave (mmWave) non-line-of-sight (NLOS) perception benchmark and open-source physical simulator for everyday objects, capturing 24 million raw radar frames and 550 high-resolution 3D SAR images across 76 objects to enable occlusion-penetrating segmentation and Sim2Real shape classification.
- MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better Machine Learning Systems
-
This paper introduces MLS-Bench, a comprehensive benchmark featuring 140 tasks across 12 core machine learning domains that isolates genuine method invention from engineering tuning through frozen evaluation harnesses, baseline-calibrated scaffolding, and strict capacity budgets, revealing that frontier AI agents still struggle significantly to discover scalable, generalizable ML algorithms.
- Mobility Helps Learning: Unsupervised Model Adaptation for Object Recognition in Dynamic Environments
-
MoCaFe introduces a model-agnostic, hyperparameter-insensitive test-time adaptation paradigm that aggregates multi-view temporal observations through inverse-variance weighting and selectively samples hard examples via prediction discrepancy, achieving up to ~20% accuracy gains without human intervention.
- Model-Agnostic FDR Control via Group Gaussian Mirror and Permutation SHAP
-
Addressing the failure of coordinate-wise FDR selection in sequential and grouped models where a single feature spans multiple lag or state sub-features, this paper introduces a model-agnostic FDR control framework that couples matrix-perturbed Group Gaussian Mirror with Permutation SHAP derivatives, delivering provable null symmetry, high statistical power, and distribution-free feature selection across linear and deep sequential architectures.
- ModelLens: Finding the Best for Your Task from Myriads of Models
-
Introduces ModelLens, the first open-world model recommendation framework that learns a latent capability atlas from 1.62M public leaderboard records, additively decoupling structural scaling priors from task-specific residual interactions to accurately rank unseen models on unseen datasets without requiring candidate forward passes.
- MOF-VERIFY: A Failure-Aware Agentic Harness for MOF Hypothesis Verification
-
To resolve verification unreliability in AI materials co-scientists caused by alias ambiguity, condition sensitivity, evidence gaps, and computation requirements, this paper builds a 4-task diagnostic benchmark (T-MOF-1–4) and proposes MOF-Verify, a failure-aware agentic harness integrating rule-based resolution, evidence sufficiency assessment, and MLIP calculation.
- MoME: Mixture-of-Memory Embeddings for Context-Aware Sparse Lookup
-
MoME (Mixture of Memory Embeddings) replaces the deterministic, single-row memory slots in existing token-indexed lookup tables with a mixture of \(M\) memory slots conditioned on the transformer's hidden state via a lightweight gate, eliminating the context-blind bottleneck of prior conditional memory while achieving consistent validation perplexity and zero-shot reasoning gains at negligible latency overhead.
- MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and Tokens
-
Presents Meta-Harness and its single-phase joint-reward formulation MoMHa for optimizing Python LLM wrapper code across accuracy, behavioral safety, and token cost, outperforming two-phase and single-objective baselines along the 3D Pareto frontier.
- MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays
-
To resolve fragmented evaluations, brittle metrics, and neglected technical confounders in high-content cellular morphology assays, MorphoHELM introduces a unified benchmark evaluating representation models across mechanism of action (MoA) enrichment, gene pathway retrieval, and replicate consistency across four progressive batch-effect tiers, demonstrating that classical CellProfiler features under standardized normalization outperform deep learning foundation models on general replicate retrieval tasks.
- MoSE3: Learning World-Space SE(3) at Every Pixel
-
MoSE3 is the first feed-forward model predicting per-pixel world-space 6-DoF rigid transforms from monocular RGB video by decomposing SE(3) motion into 3D point tracks and rigidity embeddings linked via differentiable weighted Horn fitting, trained on the new articulated benchmark Art-Kubric to achieve SOTA motion and tracking accuracy.
- Motif-Mamba: Network Motif Improved Mamba for Long-Range Sequence Modeling
-
To overcome the cross-dimensional state isolation caused by diagonal transitions in Mamba, Motif-Mamba incorporates a low-rank recurrent coupling pathway governed by three-node network motif priors, boosting long-sequence extrapolation, general language modeling, and brain–computer interface decoding while strictly preserving linear-time complexity.
- MSCR: Jointly Balancing Utilization Across Modalities and Discovering Synergistic Information
-
Addressing the issue where modality imbalance suppresses the emergence of cross-modal synergistic information, this paper proposes the Multimodal Synergistic Competition Regularizer (MSCR), which curbs unimodal dominance and explicitly rewards joint predictive gains beyond an additive unimodal baseline, achieving consistent performance improvements and dramatic gains in conditional recovery across five benchmarks.
- Multi-Part Object Representations via Graph Structures and Co-Part Discovery
-
Addressing the failure of implicit object-centric learning in occluded and out-of-distribution scenarios, this paper introduces Explicit Compositional Network (ECO-Net), which builds explicit part graphs and employs a tractable co-part discovery algorithm to compose recurring local parts into robust object representations.
- Multi-site PPG: An In-the-Wild Physiological Dataset from Emerging Multi-Site Wearables
-
The paper presents the first multi-site in-the-wild physiological dataset collected from four everyday wearable form factors (smart earring, ring, watch, and necklace) with over 350 hours of raw streams and 230–290 hours of aligned windows, benchmarking heuristic, supervised, and self-supervised heart-rate estimation across anatomical locations and multimodal fusion setups.
- Multi-Armed Bandits with Best-Action Queries
-
This paper completely resolves the theoretical value of limited best-action queries under partial bandit feedback: proving that queries cannot improve the \(\Omega(\sqrt{T-k})\) lower bound in correlated or adversarial environments, while achieving a matching optimal regret of \(\widetilde{\mathcal{O}}(\min\{Tn/k, \sqrt{n(T-k)}\})\) in stochastic i.i.d. settings via front-loaded querying paired with UCB-V.
- Multi-site PPG: An In-the-Wild Physiological Dataset from Emerging Multi-Site Wearable Devices
-
Addressing the severe wrist-centric bias and scripted laboratory constraints of existing photoplethysmography (PPG) benchmarks, this paper introduces Multi-site PPG, the first large-scale in-the-wild physiological dataset spanning four unobtrusive wearable form factors—earring, ring, watch, and necklace—and systematically benchmarks heuristic, deep supervised, self-supervised, and cross-modal fusion methods.
- Multilevel and Sequential Monte Carlo for Training-Free Diffusion Guidance
-
This paper introduces SMC-MLMC, an unbiased Sequential Monte Carlo framework with Multilevel Monte Carlo variance reduction for training-free diffusion guidance, eliminating point-estimate posterior bias while lowering cost-per-success by up to \(3\times\).
- MyoChallenge 2025: A New Benchmark for Human Athletic Intelligence
-
MyoChallenge at NeurIPS 2025 establishes a pioneering high-fidelity musculoskeletal benchmark for human athletic intelligence across table tennis rally and soccer penalty tracks, comprehensively evaluating physics-based planning, differential flatness, muscle synergies, and online behavior cloning on ultra-high-dimensional biomechanical digital twins.
- NARRA-Gym for Evaluating Interactive Narrative Agents
-
Introduces NARRA-Gym, the first executable benchmark environment for evaluating interactive narrative agents over multi-turn interaction, systematically revealing the fundamental divergence between narrative surface fluency and resistance-sensitive personalization across frontier LLMs.
- Narrowing the Collaboration Gap, Probably
-
Addressing the multi-agent decision aggregation barrier under distributed private information, this paper proves that communicating calibrated action probabilities breaks the information collapse of action-only protocols, and introduces a post-hoc conversation calibration algorithm to eliminate confidence distortions and sycophantic anchoring.
- Native Multi-Dimensional Subquadratic Operators via Input Dependent Long Convolutions
-
HyenaND proposes a natively multi-dimensional subquadratic operator that combines continuous implicit neural kernels, register-based FiLM sample-level dynamic kernel modulation, and fused nSubQ CUDA FFT routines, breaking past attention's quadratic wall and recurrent models' spatial rasterization distortions while preserving \(\mathcal{O}(L\log L)\) computational complexity.
- Native Multi-Dimensional Subquadratic Operators via Input Dependent Long Convolutions
-
HyenaND proposes a natively multi-dimensional subquadratic operator that combines continuous implicit neural kernels, register-based FiLM sample-level dynamic kernel modulation, and fused nSubQ CUDA FFT routines, breaking past attention's quadratic wall and recurrent models' spatial rasterization distortions while preserving \(\mathcal{O}(L\log L)\) computational complexity.
- Natural Jammers: When Non-Robust Features Become Antagonistic
-
This paper overturns the prevailing view that non-robust features serve purely as opportunistic shortcuts on clean data, demonstrating that they can act as antagonistic "Natural Jammers" that actively mislead standard models, driven by a spectral-spatial locking mechanism and subject to the Substitution Paradox under geometric shifts.
- Nearly Optimal Fixed-Confidence Best-Arm Identification with 1-Bit Feedback
-
Under a distribution-free finite-variance setting where rewards are only observable via 1-bit threshold queries, this paper constructs a time-uniform mean estimation primitive using a clipped tail-integral identity, introduces fixed and phased adaptive clipping algorithms, and establishes matching information-theoretic lower bounds demonstrating the inevitability of the logarithmic penalty.
- Negation Neglect: When Models Fail to Learn Negations in Training
-
This paper discovers the "Negation Neglect" phenomenon in LLM finetuning: training on documents that explicitly frame a claim as false causes models to internalize the claim as true across diverse downstream tasks, revealing a stark divide between in-weight learning and in-context reasoning that also extends to epistemic qualifiers and harmful behaviors.
- Network Intervention by Polling Strategic Agents
-
To resolve the triply entangled challenge of unobservable private technologies, strategic manipulation, and cubic computational scaling in network games, this paper develops Poll, a distributed polling algorithm based on a centrality-based decomposition of the welfare kernel that induces truthful reporting via Knightian uncertainty.
- Network of Theseus (like the ship)
-
Network of Theseus (NoT) progressively replaces layers of a guide network with target modules via representational alignment, breaking the convention that training and inference must share the same architecture and enabling radical conversions such as Conv-to-MLP and Transformer-to-RNN—even transferring structural priors from untrained guide networks.
- Neural Backward Filtering Forward Guiding
-
To overcome the intractability of Doob's \(h\)-transform in nonlinear tree-structured smoothing, this paper introduces Neural Backward Filtering Forward Guiding (NBFFG), which constructs a variational posterior by correcting closed-form linear-Gaussian proxy backward messages with neural residuals and leverages unbiased pathwise subsampling to scale training from full-tree to path-length complexity.
- Neural-Behavioral Representation of Natural Whole-body Movement in Monkeys
-
This paper presents an integrated experimental and computational neural-behavioral framework for freely moving monkeys, coupling distributed epidural cortical electrophysiology with synchronized multi-view motion capture to autoregressively decode continuous, realistic 3D whole-body movements without hand-crafted physical constraints.
- Neural Continuous-Time Markov Chain: Discrete Diffusion via Decoupled Jump Timing and Direction
-
The paper introduces Neural Continuous-Time Markov Chain (Neural CTMC), which decouples the reverse discrete diffusion rate matrix into an exit rate (when to jump) and a jump distribution (where to jump). By decomposing the path-space ELBO into a Poisson KL and a categorical KL divergence, it enables pure uniform noise diffusion to outperform prevailing masked diffusion baselines on text generation and self-correction.
- Neural-DISCO: Source-Conditioned Counterfactual Editing of Neural Population Activity
-
Neural-DISCO introduces a disentangled neural population generative framework that partitions latent representations into label-specific factors and an asymmetric noise-bottlenecked source residual, enabling precise counterfactual editing of measured behavioral/stimulus variables and single-neuron attribution while preserving unobserved source-specific brain states.
- Neural Field Thermal Tomography: A Differentiable Physics Framework for Non-Destructive Testing
-
To overcome the severe gradient decoupling and collapse of soft-constrained PINNs in stiff parabolic inverse heat conduction problems, this paper introduces NeFTY, which couples coordinate neural fields with a differentiable implicit-Euler heat solver using harmonic-mean interface fluxes and discrete adjoint gradients, achieving the first label-free 3D volumetric thermal tomography tractable on a single GPU.
- Neural Field Thermal Tomography: A Differentiable Physics Framework for Non-Destructive Testing
-
To overcome the severe gradient decoupling and collapse of soft-constrained PINNs in stiff parabolic inverse heat conduction problems, this paper introduces NeFTY, which couples coordinate neural fields with a differentiable implicit-Euler heat solver using harmonic-mean interface fluxes and discrete adjoint gradients, achieving the first label-free 3D volumetric thermal tomography tractable on a single GPU.
- Neural Galerkin Normalizing Flows for Bayesian Inference of Diffusions with Inaccessible Boundaries
-
For continuous-time stochastic diffusion processes characterized by inaccessible boundaries and vanishing diffusion matrices, this paper proposes a surrogate transition probability density function based on bounded truncated-Gaussian mixture normalizing flows and Neural Galerkin temporal projection, eliminating expensive online diffusion-bridge simulation and real-time Fokker-Planck PDE solves to enable scalable, high-accuracy MCMC Bayesian parameter inference.
- Neural networks are more modular than single neurons suggest
-
Challenging the widespread dogma that artificial neural networks lack modularity due to single-neuron mixed selectivity, this paper introduces a formally nested geometric hierarchy—spanning coordinate permutation, orthogonal rotation, and bounded invertible maps—proving via causal ablation and selective retrieval that robust functional double dissociations are preserved within multi-dimensional subspaces across vision, language, and recurrent networks.
- Neural Statistical Functions
-
The paper introduces a prefix statistics formulation and a first-order differential identity learning framework that enables neural networks to directly predict integrals, quantiles, and extreme values over arbitrary continuous operating intervals via two endpoint evaluations or logarithmic bisection without Monte Carlo sampling.
- Neural-Visual Decoding via Cognitive-guided Adaptive Blurring and Information-Constrained Alignment
-
Addressing the severe cross-modal information granularity mismatch and low SNR in EEG-based visual decoding, this paper proposes the CAIA framework, which dynamically fuses foveal center bias and saliency-guided adaptive blurring using EEG-conditioned cross-modal attention, screens task-relevant neural oscillations via information bottleneck spectral sparsity regularization, and rectifies alignment biases using distribution-aware boundary calibration to establish new state-of-the-art results on THINGS-EEG and THINGS-MEG.
- Neural-Visual Decoding via Cognitive-guided Adaptive Blurring and Information-Constrained Alignment
-
Addressing the severe cross-modal information granularity mismatch and low SNR in EEG-based visual decoding, this paper proposes the CAIA framework, which dynamically fuses foveal center bias and saliency-guided adaptive blurring using EEG-conditioned cross-modal attention, screens task-relevant neural oscillations via information bottleneck spectral sparsity regularization, and rectifies alignment biases using distribution-aware boundary calibration to establish new state-of-the-art results on THINGS-EEG and THINGS-MEG.
- Neuron Populations Exhibit Divergent Selectivity with Scale
-
This paper systematically investigates how cross-model recurring Rosetta Neurons scale with model size, revealing that their population follows a sublinear power law while occupying a shrinking fraction of total neurons, alongside a Neuron Polarization Effect (Rosetta units become increasingly monosemantic with scale while non-Rosetta units remain polysemantic) and progressive domain specialization.
- Neuronal Identity as an Organizational Basis for Analyzing Neural Population Dynamics
-
This paper proposes an identity-guided soft routing framework that groups unaligned neurons into shared latent functional groups using their multimodal biological profiles, achieving state-of-the-art zero-shot cross-subject neural decoding and manifold geometry preservation without retraining or neuron-wise correspondence.
Neuronal Identity as an Organizational Basis for Analyzing Neural Population Dynamics
- Neuronal Self-Adaptation Enhances Capacity and Robustness of Representation in Spiking Neural Networks
-
Inspired by biological voltage-gated potassium channel self-regulation, this paper proposes the Potassium-regulated Leaky Integrate-and-Fire (KvLIF) neuron model, which introduces an auxiliary conductance state to dynamically execute subthreshold synaptic shunting and post-spike adaptive afterhyperpolarization reset, substantially extending the dynamic response range and enhancing noise robustness in SNNs.
- NIV: Neural Axis Variations for Variable Font Generation
-
NIV introduces the first neural geometric deformation framework that automatically converts a single static vector font into a fully functional OpenType variable font (.ttf), predicting contour control point displacements via sequence-to-sequence attention and a novel Property Embedding mechanism, coupled with a hierarchical residual tuple construction engine for continuous multi-axis interpolation.
- Normative Networks for Source Separation via Local Plasticity and Dendritic Computation
-
The paper introduces Predictive Entropy Maximization (PEM), deriving a biologically plausible online neural network for blind source separation by applying a second-order Taylor expansion to regularized log-determinant output entropy, thereby enabling local dendritic error-driven plasticity and local Hebbian lateral inhibition without non-local inverse-covariance calculations.
- Not Too Generative, Not Too Discriminative: The Human Alignment Sweet Spot
-
This paper uses Joint Energy-Based Models (JEMs) to continuously interpolate between discriminative and generative learning objectives within a strictly fixed network architecture, systematically demonstrating across six benchmarks that human visual alignment consistently peaks in an intermediate hybrid sweet spot rather than at either extreme.
- Occupancy-based Quantile Risk Control
-
This paper introduces Occupancy-based Quantile Risk Control (OQRC), reformulating quantile-based risk control as a finite-occupancy problem over partitioned calibration losses, establishing provably tight risk bounds with finite-sample validity and reducing the risk gap by up to 78.64%.
- OFBD: Object-Focused Background Debiasing for Long-Tailed Learning
-
Addressing the core finding that tail-class degradation in long-tailed visual recognition stems heavily from spurious background shortcuts in both representation and optimization dynamics, this paper proposes OFBD (Object-Focused Background Debiasing), which combines RL-based foreground-adaptive CutMix (FG-CutMix) and parameter-free channel-statistical background feature rectification (BFR) to substantially boost tail-class performance without external data or pretrained weights.
- Offline Materials Optimization with CliqueFlowmer
-
Addressing the fundamental exploration limits of maximum-likelihood generative models in computational materials discovery, CliqueFlowmer maps hybrid discrete-continuous crystal structures into a fixed-dimensional latent space endowed with a clique decomposition, enabling direct target property optimization via derivative-free evolution strategies and continuous flow decoding.
- Offline Preference-Based Trajectory Evaluation
-
The paper introduces an offline preference-based trajectory evaluation framework for agentic systems that directly compares multi-step trajectories via non-parametric temporal preferences over progress and time-to-return profiles, reducing instance-level ties from roughly 75% to 35% and substantially boosting discriminative power, ranking stability, and data efficiency.
- Omni-Interactive Universal Embedder
-
OmniUE breaks beyond the traditional text-only interaction paradigm by integrating off-the-shelf SAM-3 visual and SAM-Audio acoustic segmenters with learnable sequence tokens and cross-layer adaptive importance aggregation, establishing the first universal embedder capable of processing arbitrary multimodal inputs (text, silent video, audio) conditioned on omni-interactive prompts (text, visual ROI regions, and audio temporal spans).
- OmniShotCut: Holistic Relational Shot Boundary Detection with Shot-Query Transformer
-
OmniShotCut reformulates shot boundary detection as structured relational prediction, leveraging a shot query-based dense video Transformer to jointly estimate shot temporal boundaries, intra-shot transition classes, and inter-shot continuity relationships, supported by a scalable self-supervised synthetic transition pipeline.
- On Observation Time for Recovering Latent Hawkes Networks
-
From an information-theoretic statistical perspective, this paper proves that exact topology recovery of sparse, weakly-coupled stationary Hawkes networks requires an optimal continuous observation time scaling logarithmically with network size, \(T \asymp \log d\), verified through a matching two-stage screening-and-regression estimator and Fano lower bounds.
- On the Design Space of Discrete Diffusion Online Adaptation for Molecular Optimization
-
This paper systematically deconstructs the design space of online test-time adaptation for pretrained discrete diffusion models in molecular optimization, showing that Thompson sampling, CVaR reward shaping, and single-pass density entropy debiasing synergize to achieve up to a 58x sampling acceleration over inference-time search under matched oracle budgets.
- On the Error Correcting Effects of Stochasticity in Discrete Diffusion
-
The paper formally proves the error-contracting and error-correcting mechanisms of stochastic transitions in discrete diffusion through the lenses of information theory and total variation distance, and introduces Discrete Churn and Restart Sampling (DCRS)—an inference algorithm alternating forward noise perturbations with reverse near-deterministic steps—achieving up to a 10x reduction in sampling steps under the low number of function evaluations (NFE) regime.
- On the Meta-Design of Allocation Problems
-
This paper introduces a formal meta-design framework for societal resource allocation problems, constructing empirical "welfare surfaces" over predictive data collection, capacity constraints, and intervention size to guide Pareto-optimal budget investments before downstream policy optimization.
- On the Meta-Design of Allocation Problems
-
Departing from the conventional policy learning paradigm that treats data collection, program capacity, and intervention quality as fixed constraints, this paper formalizes the meta-design space of resource allocation problems and introduces a geometric framework based on the welfare surface to empirically evaluate and optimize trade-offs between prediction investments and capacity expansions across real-world social assistance domains.
- On the Recoverability of Causal Relations from Bulk Gene Expression Data
-
Background and Motivation¶
- One-Shot Private Confidence Regions via Resampling
-
Addressing the issue where traditional differentially private resampling incurs a privacy loss scaling as \(\sqrt{B}\) due to privatizing each intermediate resample statistic, this paper proposes a one-shot noise addition framework that perturbs only the final resampling quantile radius, reducing the privacy dependence on the number of resamples \(B\) to logarithmic or entirely independent.
- One Model, Two Roles: Emergent Specialization in a Shared Recurrent Transformer
-
Introducing Asymmetric Input Recurrence (AIR), this paper demonstrates that a single shared-weight Transformer without physical modularity spontaneously develops distinct computational roles—a committed proposal state and an uncertain scratchpad—induced solely by asymmetric input injection or a separable level token.
- Online Allocation with Unknown Shared Supply
-
For multi-site online prepositioning under unknown shared supply and fixed-charge shipping costs, this paper introduces a deterministic threshold-proportional algorithm GPA alongside a learning-augmented extension, establishing a tight 4/3 competitive ratio and Pareto-optimal prediction guarantees.
- Online Bayesian Calibration under Gradual and Abrupt System Changes
-
To address gradual drift and abrupt regime shifts in nonstationary streaming data, this paper proposes Bayesian Recursive Projected Calibration (BRPC), which decouples discrepancy-free parameter particle updates from conditional Gaussian process discrepancy updates to preserve parameter identifiability, augmented with change-point restart mechanisms for robust online adaptation.
- Optimal and Efficient Contextual Combinatorial Semi-bandits with General Function Approximation
-
For contextual combinatorial semi-bandits with general function approximation, this paper proposes SquareCB.Comb, the first computationally efficient algorithm achieving the minimax optimal regret bound of \(\widetilde{\mathcal{O}}(\sqrt{mAT\log|\mathcal{F}|})\) by solving a log-barrier regularized convex program over base-arm participation vectors.
- Optimal Ansatz-Free Hamiltonian Learning in Situ
-
Addressing the hardware bottleneck of lacking coherent control and ancillary qubits on near-term quantum platforms, this paper proposes the first control-free, ancilla-free ansatz-free Hamiltonian learning algorithm that uses only Pauli product state preparation and measurement, achieving the optimal SQL total evolution time \(\Theta(\frac{\Lambda}{\varepsilon^2}\log\frac{\Lambda}{\varepsilon})\).
- Optimal Sequential Tests Yield Log-Optimal e-Processes
-
Addressing arbitrary composite null hypotheses, this paper proves that a family of asymptotically optimal sequential tests can be weighted and aggregated into an asymptotically log-optimal e-process (the WAIT process), characterizes the deterministic log-growth rate via a profile exponent, and clarifies the structural discrepancies across various test optimality criteria.
- Optimal Transport Reweighting for Robust Learning under Spurious Correlations and Label Noise
-
POTER introduces an optimal transport reweighting framework that derives sample importance from class-conditioned dual potentials against a validation-constructed reference distribution, simultaneously resolving spurious correlation shortcuts and label noise within a single standard ERM stage without retraining.
- OrthoReg: Orthogonal Regularization for Hybrid Symbolic-Neural Dynamical Systems
-
Addressing the overlap pathology where neural augmentations relearn mechanistic terms under sparse symbolic discovery due to continuous shrinkage bias, OrthoReg directly penalizes their empirical inner product, ensuring a complementary decomposition that preserves interpretability and boosts out-of-distribution generalization.
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning
-
PaLoRA reveals finite-precision subspace leakage in LoRA-based continual learning, derives an optimal rank-aware pacing law \(s^* = \sqrt{R/c}\) for gradient scaling, and couples it with adaptive SVD compression and nullspace projection to achieve remarkable stability on long-horizon continual benchmarks.
- PaperScorer: How Predictable Is Paper Acceptance?
-
PaperScorer introduces balanced, anonymized corpora that rigorously strip away formatting shortcuts and author identities, training open-source text and vision backbones directly on binary peer-review decisions; results demonstrate that vision models utilizing full-page screenshots significantly outperform frontier reasoning systems, achieving calibrated acceptance probabilities that effectively anchor downstream retrieval and agentic reviewing.
- Parabolic Position Encoding: Vision-Centric, Principled, Extrapolatable, General
-
Addressing the mismatch between language-derived position encodings and multi-dimensional vision characteristics, this paper proposes Parabolic Position Encoding (PaPE), which decomposes attention bias into a sum of learnable parabolas over relative token coordinates to simultaneously achieve translation invariance, directionality, distance decay, and context awareness, outperforming the next-best method by 10.5% in zero-shot classification extrapolation at \(1024^2\) resolution while natively supporting FlashAttention kernels.
- Paradoxes of Game Theoretic Equilibria and Price of Anarchy
-
This paper systematically reveals the foundational topological failure of reducing multi-agent learning to static equilibrium concepts (NE/CE/CCE) and the Price of Anarchy (PoA), proving that worst-case pure Nash equilibria are dynamically unstable strict saddles or global repellers, that PoA becomes unbounded under data-driven affine costs, and that discrete-time learning in non-atomic routing natively induces Li-Yorke chaos with an exponential \(2^p\) degradation in social welfare.
- ParaLin: Accelerating Parallel Diffusion Integrator via Intrinsic Partially Linear Decomposition
-
Addressing the severe quality degradation of conventional few-step samplers on dense instructions and in-image text rendering, this paper proposes ParaLin++, a step-parity parallel diffusion integrator that leverages intrinsic partially linear decomposition, successive over-relaxation, and adaptive stopping to achieve 1.68×–4.8× wall-clock acceleration while strictly preserving full-step generation quality.
- Partition Tree: Conditional Density Estimation over General Outcome Spaces
-
This paper introduces Partition Tree and its bagging extension Partition Forest, which model conditional distributions as piecewise-constant densities on data-adaptive joint partitions using Radon–Nikodym derivatives, optimizing conditional negative log-likelihood across continuous, categorical, and mixed outcomes.
- Path Dependence under Adaptive AI Delegation
-
This paper develops a non-linear dynamical system framework modeling long-run human-AI interaction, proving that performance-driven adaptive AI delegation induces bistability and an interior saddle separatrix, causing subtle initial skill or reliance differences to trigger path-dependent divergence and cognitive debt.
- PCFBench: How Far Can Large Vision-Language Models Go in Physics-aware Photonic Inverse Design?
-
To tackle the challenge of tightly coupled geometric intuition and deep waveguide physics in photonic crystal fiber (PCF) inverse design, this paper introduces PCFBench—the first physics-grounded benchmark covering the perception-understanding-reasoning-decision pipeline (comprising 18 fiber families, 220K FDTD simulation samples, and 26 hierarchical tasks)—and evaluates 41 frontier VLMs alongside 13 multi-agent systems, exposing a stark perception–design gap and severe error cascade dynamics.
- Peer review should constrain evaluative authority
-
This position paper argues that because scientific peer review operates under persistent evaluative uncertainty and missing criteria ground truth, institutional reform should shift from improving evaluators or aggregating opinions to constraining evaluative authority before consequential judgments become decision-relevant, demonstrating the approach through a dual-track human–AI mutual supervision framework.
- PENEX: AdaBoost-Inspired Neural Network Regularization
-
To resolve the optimization divergence and constrained formulation challenges of multi-class exponential loss in deep learning, PENEX introduces a SumExp soft penalty barrier with closed-form optimal parameter estimation, enabling first-order gradient descent to explicitly maximize classification margins and resist overfitting.
- Perturb and Correct: Post-Hoc Ensembles using Affine Redundancy
-
Addressing the failure of single-model representations under distribution shift and the prohibitive cost of Deep Ensembles, Perturb-and-Correct (P&C) pairs random hidden-layer parameter perturbations with closed-form least-squares corrections in following affine layers, preserving in-distribution behavior while exposing structured epistemic disagreement on out-of-distribution data.
- PiCA: Pivot-Based Credit Assignment for Search Agentic Reinforcement Learning
-
PiCA addresses reward sparsity, isolated credit assignment, and distributional shift in search agent reinforcement learning by modeling trajectories as cumulative success gains under potential-based reward shaping, establishing new state-of-the-art results across seven QA benchmarks.
- Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation
-
This paper demonstrates that goal-directed navigation planning naturally emerges from the physical relaxation dynamics of a hippocampal recurrent network clamped at a target goal, proving exact mathematical equivalence to the desirability field of a Linearly-solvable Markov Decision Process (LMDP) and enabling immediate low-rank global replanning upon local barrier edits.
- Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry
-
This paper extends the Strong Platonic Representation Hypothesis to biological neural systems, demonstrating that subject-specific fMRI representations can be translated across human brains via unsupervised orthogonal rotations and global synchronization without paired stimuli or intermediate models.
- Platonic Task Arithmetic
-
Based on the Platonic Representation Hypothesis, this paper proposes Universal Task Descriptors (UTDs) that transcend weight coordinates, enabling cross-architecture task addition and selective task negation across heterogeneous models via closed-form linear operator folding or low-rank adapters.
- Playing ZendoWorld: Challenging AI Agents on Active Visual Concept Induction
-
This paper introduces ZendoWorld, an interactive benchmark where agents perceive 3D geometric scenes, actively design new experiments to test hypotheses, and induce underlying logical rules, demonstrating that high labeling accuracy does not imply rule recovery and that current VLMs generate largely uninformative experiments compared to humans.
- PMO-Dock: Benchmarking Docking, Specificity, and Generalization in Molecular Optimization
-
PMO-Dock establishes a standardized benchmark of 25 docking-based molecular optimization tasks that strictly separates hyperparameter development from held-out test evaluation and introduces a novel binding specificity objective, revealing that no single optimization algorithm dominates across drug design workflows and that hyperparameter transfer across targets is highly fragile.
- PolarScale: A Physics-Grounded Benchmark for Radiometrically Consistent RGB-to-Stokes Estimation
-
Addressing the issue that existing RGB-to-polarization methods normalize out the radiometric scale by dividing by scene maximum intensity, PolarScale introduces a benchmark making per-scene radiometric scale \(\sigma\) an explicit prediction and evaluation target, accompanied by constant-scale controls, wrap-aware angular metrics, self-consistency, and physical-bound checks.
- Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses
-
Grounding AI political neutrality in political theory and conflict mediation, this paper proposes the concept of Maximum Equal Approval (MEA) on the Pareto frontier and demonstrates through the PARETO benchmark (7,434 participants, 208,152 evaluations) that balanced AI responses can achieve broad bilateral approval with less than a 10% penalty compared to one-sided agreement.
- Position: Let’s Strengthen Verifiability If We Can’t Enforce Reproducibility
-
Drawing on a code-availability survey of five leading ML/CV conferences from 2021–2025, this paper proposes integrating experiment-log checks and metric recomputation from prediction files into submission, review, and publication when full reproduction cannot be enforced; it is a proposal for partial consistency verification, not a demonstrated fraud-detection system.
- Post-Selection Distributional Model Evaluation
-
Addressing the optimistic post-selection bias when candidate ML models are evaluated on the same data used for pre-screening, this paper introduces PS-DME, an e-value inversion framework that constructs valid CDF confidence bands under full data reuse while controlling the false coverage rate (FCR).
- Posterior-First Neural PDE Simulation: Inferring Hidden Problem State from a Single Field
-
Addressing the ill-posed nature of predicting futures from a single observed field, this paper establishes a "posterior-first" neural PDE simulation framework that infers a posterior over task-sufficient hidden problem states before conditional rollout, proving an irreducible ambiguity barrier for point estimates and substantially closing the gap to oracle metadata on both synthetic and PDEBench benchmarks.
- PosteriorBench: From Point Estimates to Posterior Matching in Evaluating Generative Inverse Solvers
-
Addressing the fundamental flaw where point estimate benchmarks fail to capture ill-posed multi-solution behavior in scientific inverse problems, this paper introduces PosteriorBench—the first distribution-centric benchmark featuring high-fidelity reference posteriors and a five-metric evaluation suite across four physical domains.
- PreCoMem: Predictive Cognitive Memory for Self-Evolving Long-Term Dialogue Agents
-
Inspired by predictive coding theory, PreCoMem represents long-term memory as confidence-weighted beliefs, using reliability-weighted Effective Surprisal and a two-threshold gate to dynamically schedule reinforcement, hypothesis buffering, and soft-decay correction, achieving robust adaptation to non-stationary belief drift while keeping effective memory footprint bounded.
- Prediction-Powered Active Testing
-
This paper introduces Prediction-Powered Active Testing (PPAT), a label-efficient evaluation framework that combines the unbiased LURE estimator with cheap black-box control variates to residualise the loss and reduce sampling variance, while establishing the first asymptotically valid confidence intervals for active testing.
- Predictive Geometry of Hidden Trajectories in Transformers
-
By pulling the terminal prediction loss back to intermediate residual states through the fixed downstream computation, this paper derives the pullback Fisher curvature operator and causal token sensitivity scores, revealing severe predictive anisotropy in hidden trajectories and guiding non-uniform rank allocation, token pruning, and low-rank distillation.
- PRIM: Meta-Learned Bayesian Root Cause Analysis
-
The paper introduces PRIM, a graph-free root cause analysis method based on prior-fitted transformer neural processes that marginalises out graph structure and mechanism uncertainties via Bayesian model averaging, achieving zero-shot root cause identification in 17 ms per forward pass without requiring prior causal graphs or test-time graph discovery.
- Primal-Dual Guided Decoding for Constrained Discrete Diffusion
-
Addressing the heavy computational overhead and auxiliary classifier dependencies of existing constrained discrete diffusion decoding, this paper introduces Stochastic Primal-Dual guided Decoding (SPDD), which reformulates constrained generation as online KL-regularized optimization via mirror descent to adaptively adjust token logits with minimal inference overhead and superior distributional fidelity.
- Principia: Relational Physics Tests for Video Models
-
Principia introduces the first calibration-independent physical reasoning benchmark for video models based on intra-scene paired-object relational invariants across eight Newtonian phenomena, demonstrating that visual realism (VBench ~0.8) and physical consistency (scores under 0.42) are completely decoupled in modern video generators.
- PRISM: Prior Relational Information for Self-Supervised Modeling to Enhance Solubility Prediction
-
PRISM is a label-free molecular pretraining framework that converts expert-defined chemical priors into relational teachers via soft neighborhood distribution matching across four decoupled encoder branches, combined with adaptive late fusion and composition-weighted heads to dramatically enhance solubility out-of-distribution (OOD) generalization across unseen scaffolds, solutes, solvents, and solvent mixtures.
- Privacy by Postprocessing the Discrete Laplace Mechanism
-
This paper introduces a deterministic post-processing unbiased estimator for nonlinear functions evaluated on data privatized by the discrete Laplace mechanism, proves its uniqueness and minimal variance among all unbiased estimators, and demonstrates that the discrete Laplace mechanism can be transformed into continuous Laplace or Staircase mechanisms via independent noise post-processing.
- Private Online Prediction from Experts with Small Losses
-
This paper presents the first differentially private algorithm for online prediction from experts that attains a data-dependent low-loss regret bound of \(\tilde{O}(\sqrt{L^\star})\) against oblivious adversaries, nearly matching non-private lower bounds for constant privacy parameters, while establishing that an \(\tilde{\Omega}(\sqrt{T})\) regret is unavoidable against adaptive adversaries even when \(L^\star = 0\).
- Privately Estimating Monotone Statistics in Polynomial Time
-
This paper investigates black-box differentially private estimation of monotone statistics, establishing an interleaving property of subsampled empirical quantiles to design polynomial-time algorithms that reduce sample complexity by a factor of \(t\) at the cost of \(e^{O(t)}\mathrm{poly}(k_{\varepsilon,\delta})\) runtime, alongside a matching exponential query complexity lower bound.
- ProCARE: Real-World Study Automation via Profile-Grounded Evidence Contracts
-
Addressing frequent schema mismatches, temporal inversions, and unsupported claims in LLM-agent-driven real-world clinical study automation, ProCARE reframes the process as constrained compilation via profile-grounded evidence contracts, executing strongly typed RWS-IR, localized violation repair, and certified claim admission.
- Profit Maximization in Bilateral Trade against a Smooth Adversary
-
Addressing the non-parametric online profit maximization problem in bilateral trade, this paper proposes a hierarchical learning algorithm Hier-Mech based on algorithmic chaining, achieving a minimax-tight \(\tilde{O}(\frac{1}{\sigma}\sqrt{T})\) regret bound against a \(\sigma\)-smooth adversary and establishing a formal reduction to the joint ads problem.
- Program-as-Weights: A Programming Paradigm for Fuzzy Functions
-
This paper introduces the Program-as-Weights (PAW) programming paradigm, which uses a 4B neural compiler to compile natural-language specifications once into compact LoRA weights and pseudo-programs that run on a frozen 0.6B local interpreter at 30 tokens/s offline, matching Qwen3-32B prompting accuracy at approximately one-fiftieth of the memory footprint.
- Project Auto-World: Towards Automated Benchmarking of Neural Relational Reasoners
-
Project Auto-World introduces an end-to-end framework combining LLM-driven evolutionary program search (FunSearch) and autonomous research agents (Auto-Research) to synthesize challenging knowledge graph reasoning instances under 8 entities, uncovering blind spots in state-of-the-art Edge Transformers that evade traditional difficulty metrics.
- Propagation of Chaos in Contextual Flow Maps
-
By formalizing deep transformers as contextual flow maps where a distinguished token evolves alongside an empirical contextual measure, this paper establishes uniform forward depth bounds and uniform backward online gradient descent training trajectory bounds converging to the infinite-context population limit, proving a dimension-free parametric rate of \(\mathcal{O}(n^{-1/2})\) via a novel Eulerian adjoint system.
- Proportionality in Ranking Compression
-
Addressing the challenge of compressing vast annotator preferences into a compact reward ensemble for pluralistic AI alignment, this paper formalizes the ranking compression problem, proves that positional proportionality and proportionality for solid coalitions (PSC) can always be computed in polynomial time, and characterizes the general impossibility and structured feasibility of pairwise proportionality.
- Provable Speedups From Dynamic Population Sizes in Evolutionary Algorithms for Multiobjective Optimization
-
This paper theoretically demonstrates for the first time that dynamic population sizes provably accelerate evolutionary multiobjective algorithms: on the constructed bi-objective benchmark CLIMB, NSGA-II-DYN with dynamic population adjustment and GSEMO cover the entire Pareto front in expected \(O(n \log n)\) fitness evaluations, achieving an \(\Omega(\sqrt{n}/\log n)\) super-constant speedup over the \(\Omega(n^{1.5})\) lower bound of static NSGA-II.
- Proxy-Based Approximation of Shapley and Banzhaf Interactions
-
ProxySHAP introduces an exact polynomial-time closed-form interaction extraction algorithm for tree ensembles combined with situational residual correction, achieving state-of-the-art approximation quality for higher-order Shapley and Banzhaf interactions across both low- and high-budget regimes.
- RABBIT: Rapidly Adaptive BOLD Foundation Model via Brain-Tuning for Accurate Zero-Shot Decoding
-
This paper introduces RABBiT, a compact audio-to-fMRI foundation model that pairs a brain-tuned speech backbone with a region-attentive Temporal Brain Transformer (TBT) and a low-rank Shared–Idiosyncratic Decomposition (SID), achieving state-of-the-art zero-shot fMRI prediction across 324 unseen listeners while enabling highly accurate few-shot subject calibration from just 10 minutes of fMRI data.
- RADAR: Relative Angular Divergence Across Representations
-
RADAR proposes a geometrically grounded transferability metric that evaluates layer-wise representation displacement trajectories through angular alignment and relative distance distributions, leveraging Gaussian Mixture Models and symmetric KL divergence to accurately predict cross-domain transfer gains without fine-tuning.
- RamanBench: A Large-Scale Benchmark for Machine Learning on Raman Spectroscopy
-
This paper introduces RamanBench, the first large-scale, fully reproducible benchmark unifying 77 datasets, 169 prediction targets, and over 325k spectra across four domains, showing under standardized evaluation of 28 models that tabular foundation models consistently outperform domain-specific architectures while significant headroom remains.
- Random Matrix Theory of Early-Stopped Gradient Flow: A Transient BBP Scenario
-
This paper analyzes gradient flow dynamics in a linear teacher-student setting with two-block input covariance anisotropy, establishing an exact \(2\times 2\) Dyson equation and rank-two determinant condition to analytically characterize a transient BBP transition that acts as a solvable spectral mechanism for early stopping.
- RANSAC Scoring Done Right
-
By analytically marginalizing the noise scale under a conjugate Inverse-Gamma prior for a fixed inlier partition and optimizing via sort-and-sweep, this paper derives the first genuinely scale-free RANSAC score computable in \(\mathcal{O}(N \log N)\) time.
- Re-evaluating Continual Learning with Few-Shot Adaptation
-
This work challenges the standard continual learning evaluation paradigm of measuring stability strictly by zero-shot recall and plasticity by current-task accuracy, demonstrating through few-shot test-time adaptation that catastrophic forgetting is largely recoverable, while proposing foresight meta-learning and the SAUCE metric to capture rapid learning-to-learn dynamics.
- Read, Parse, Describe: Unified Document Parsing with Visual Element Description Generation
-
This paper introduces Read-Parse-Describe, a unified document parsing benchmark requiring joint prediction of bounding boxes, reading order, text/HTML content, and accessibility alt-text, alongside SARA, a plug-and-play spatial alignment residual adapter that injects multi-layer visual cues into frozen VLMs.
- ReCoG: Relational Concept Graph for Training-Free Personalization
-
Addressing the issue where conventional training-free personalization describes concepts in isolation and suffers under fine-grained confusion, ReCoG proposes a Relational Concept Graph that explicitly elicits pairwise discriminative edges between concepts offline, followed by round-robin pairwise comparisons conditioned on relational edges at inference time, achieving substantial accuracy gains without any fine-tuning.
- Recon: Reconstruction-Guided Reasoning Synthesis for User Modeling
-
Addressing the issue that conditioning on observed actions yields post-hoc rationalizations rather than true causal decision processes in open-ended user modeling, Recon evaluates candidate reasoning traces by their action reconstruction fidelity, providing strong training-free selection and RL rewards that substantially improve user action prediction.
- Reconciling Causality and Non-Equilibrium Thermodynamics with Hamiltonian Causal Models
-
Introduces Hamiltonian Causal Models (HCMs) coupling microscopic Hamiltonian mechanics with local thermal environments, demonstrating that thermodynamic entropy production serves as a principled causal observable detecting path-level causal effects and cyclic parenthood that evade conventional average treatment effects.
- Refinement as a Service: Algorithmic Predictor Refinement
-
Addressing calibrated predictor aggregation without access to true Bayes probabilities, this paper models predictors as signaling schemes under feature-independent garblings, establishes that deterministic refinement is polynomial-time solvable bilaterally but NP-hard multilaterally, and presents a polynomial-time extreme-ray decomposition algorithm for arbitrary input sizes when randomized outputs are permitted.
- Reinforced Fast Weights via Next-Sequence Prediction
-
Addressing the mismatch between evolving recurrent memory in fast weight architectures and single-token supervision from next-token prediction (NTP), this paper proposes ReFINE, a next-sequence prediction (NSP) reinforcement learning framework that samples high-entropy positions for multi-token rollouts and optimizes them via GRPO to eliminate long-horizon memory drift.
- ReSFlow: Rectifying Categorical Flows on Statistical Manifolds for One-Step Generation
-
Addressing the computational bottleneck in categorical flow matching caused by intersecting, curved Fisher-Rao geodesics requiring expensive multi-step numerical integration, ReSFlow maps discrete distributions to the positive orthant of the unit hypersphere via a square-root diffeomorphism, rectifies trajectory intersections via iterative ODE recoupling, and achieves high-fidelity one-step discrete generation via the closed-form Riemannian exponential map.
- Rethinking Contrastive Loss in CLIP Post-training: A Complementary Framework with Frozen Text Encoder
-
Uncovering that catastrophic forgetting in contrastive post-training stems from an overly large temperature rather than insufficient negatives, this work introduces ComCLIP—a lightweight post-training recipe with a frozen text encoder that combines small-temperature contrastive loss, MSE geometric anchoring, and DINOv2 relational distillation to boost visual feature transferability while preserving drop-in compatibility for downstream VLMs.
- Rethinking Expressivity and Efficiency in Test-Time Training
-
Addressing the core dilemma between expressivity and hardware efficiency in Test-Time Training (TTT), E2-TTT derives closed-form scalar kernels under chunk-start gradient evaluation to exactly reproduce token-wise momentum and decay dynamics in a single backward pass, achieving 8x length extrapolation while matching chunk-level training throughput.
- Rethinking Learning from Label Proportions via Moment Matching
-
Addressing the vulnerability of learning from label proportions (LLP) under non-random structured bag grouping, this paper reveals that bag-level label counts follow a Poisson multinomial distribution (PMD) and introduces LLP-MM, an efficient multi-order factorial moment matching framework that bypasses intractable likelihood calculation while enforcing progressively tighter solution space constraints.
- Rethinking State Tracking in Recurrent Models Through Error Control Dynamics
-
This paper investigates recurrent state tracking through error control dynamics, proving that affine recurrent models (including SSMs and Linear Attention) inevitably act as the identity on state-separating subspaces once preserving state representations, and establishing a quantitative distinguishability ratio framework that accurately predicts downstream length generalization collapse.
- Rethinking XAI Evaluation: A Human-Centered Audit of Shapley Benchmarks in High-Stakes Decision Support
-
Using an amortized neural surrogate framework to eliminate algorithmic confounders, this paper audits eight Shapley formulations across 3,735 expert fraud case reviews, demonstrating that offline quantitative evaluation proxies are fundamentally decoupled from human utility while explanations consistently inflate decision confidence without improving objective accuracy.
- Retrieval Heads Meet Vision: Uncovering How VLMs Locate and Extract Visual Information
-
This paper uncovers the internal mechanism of visual retrieval in vision-language models (VLMs), demonstrating that a sparse subset of attention heads (~1.7%–2.6%), termed Visual Retrieval Heads (VRHs), is causally responsible for binding referring expressions to image regions, functioning as an architecturally shared circuit that generalizes across diverse visual reasoning benchmarks.
- Revealing Epistemic Uncertainty in MLLMs via Causal-Invariant Masking
-
Addressing the issue that current MLLM uncertainty quantification skews toward aleatoric noise and overlooks model-intrinsic limitations, this paper proposes the Causal-Invariant Masking (CIM) framework, which quantifies epistemic uncertainty via Semantic Divergence (SD) and introduces Expected Embedding Drift (EED) as an efficient geometric proxy.
- Reviewer Scores Are Not Comparable Across Research Areas in ML Peer Review
-
Using a large-scale empirical audit across 50,289 ICLR 2021–2026 papers and 219 research topics, this work demonstrates that research topic significantly predicts paper acceptance even after controlling for reviewer scores and year (\(p = 5.1 \times 10^{-3}\)), with extreme topic acceptance ratios reaching 8× at identical scores, establishing that recorded numerical ratings lack a unified decision meaning across areas and diagnosing a systemic measurement design failure.
- RigidFormer: Learning Rigid Dynamics using Transformers
-
Overcoming the computational bottleneck and topology dependencies of mesh- and vertex-based neural simulators, RigidFormer introduces an object-centric mesh-free Transformer simulator that integrates Anchor-based RoPE (ARoPE), Anchor-Vertex Pooling (AVP), and differentiable Kabsch rigid projection, achieving high-fidelity, time-step-controllable multi-rigid contact dynamics from point clouds at real-time speed (23.9 FPS).
- Risk-Averse Online POMDP Planning via CVaR of the Immediate Cost with Performance Guarantees
-
Addressing the safety vulnerabilities where standard POMDP expected cost planning masks catastrophic tail states in the belief and existing risk-averse methods break standard MDP dynamic programming, this paper applies Conditional Value-at-Risk (CVaR) directly to the immediate cost over the belief, simultaneously suppressing current hidden state risk, preserving standard MDP value structure, and enabling any off-the-shelf expectation-based planner to become risk-sensitive with finite-time estimation error bounds entirely independent of the risk level \(\alpha\).
- Risk-Controlled Post-Processing of Decision Policies
-
Addressing the challenge that legacy decision policies often fail strict safety risk budgets, this paper proposes a calibration-based post-processing algorithm that scores conditional violation risk reduction and switches to a fallback policy only on high-risk contexts, achieving guaranteed risk control while maximally preserving agreement with the baseline.
- Robust Domain Generalization under Divergent Marginal and Conditional Distributions
-
Addressing the compounded distribution shifts where both marginal label distributions \(P(Y)\) and class-conditional distributions \(P(X \mid Y)\) diverge across domains, this paper presents RC-Align, a meta-learning framework guided by a novel generalization bound that strictly upper-bounds feature mismatch via contrastive domain-class alignment (DA) loss, establishing new state-of-the-art performance on standard DG and multi-domain long-tailed (MDLT) benchmarks.
- Robust Multi-view Clustering against Imperfect Information
-
Addressing the co-existence of incomplete views and noisy correspondences in multi-view clustering, this paper proposes Posterior-guided Latent Counterpart Inference (PLCI), which formulates target-view counterparts as latent variables and infers them via instance-level reliability and prototype-level semantic optimal transport.
- RoSA: Rotational Sparse Adaptation for Memory-Efficient Fine-Tuning
-
RoSA introduces a depth-structured fine-tuning framework that freezes an early prefix, maintains a permanent output anchor, and rotates a single active layer block across depth with intra-block sparse optimization, enabling near-full fine-tuning accuracy within 40 GB GPUs while truncating backpropagation paths.
- RotMoLE: Enhancing Mixture of Low-Rank Experts through Rotational Gating Mechanism
-
Addressing the fundamental limitation that conventional MoE-LoRA scalar routers only rescale expert magnitudes along fixed directions, RotMoLE exploits the low-rank bottleneck to introduce lightweight rotational gating and subspace projection, enabling orthogonal 2D feature rotations that substantially expand expert capacity and multi-task fine-tuning performance under constrained expert budgets.
- RoutingBench: Can Agentic Routing Analysis Scale to Production Datacenter Networks?
-
Addressing the memory explosion and hours-long latency of traditional symbolic network analysis tools in hyperscale networks, this paper introduces RoutingBench and demonstrates that an Agent Skills paradigm guided by the "explore more; digest less" principle scales routing-path analysis to 50,000 routers with 99.5% accuracy.
- RT-J: Large-Scale Pretraining of Relational Transformers for Context-Efficient Predictions
-
RT-J introduces an 86M-parameter relational foundation model for multi-table relational databases, pretrained on the Join (639 web-crawled real-world databases and 6K forecasting tasks) with a lightweight random-walk context retriever and test-time context scaling, matching or outperforming strong feature-engineering pipelines with \(10\times\) fewer in-context examples.
- RuleSmith: Multi-Agent LLMs for Automated Game Balancing
-
RuleSmith frames asymmetric game balancing as black-box optimization over rule parameters, combining zero-shot multi-agent LLM self-play evaluations with Expected-Improvement-driven adaptive Bayesian optimization without requiring task-specific policy training.
- Runtime Monitoring of Perception-Based Autonomous Systems via Embedding Temporal Logic
-
Embedding Temporal Logic (ETL) formalizes and monitors temporal specifications directly within the latent space of pretrained visual encoders, using conformal prediction calibration to deliver distribution-free recall guarantees without requiring object detectors or symbolic labels.
- RustMizan: A Compilable, Contamination-Aware Benchmarking Framework for Rust Vulnerabilities
-
Addressing the critical limitations of existing vulnerability benchmarks—lack of compilability, trivial binary-only evaluation, and training corpus contamination—RustMizan introduces the first multi-level compilable benchmarking framework for Rust vulnerabilities across crate, file, and function scopes, paired with an automated compiler-driven mutation pipeline to probe genuine reasoning and adversarial robustness.
- Safeguarding Mutual Correction in Source-Free Domain Adaptation via Cut Statistics
-
Addressing error accumulation and confirmation bias in unsupervised domain adaptation, SafeCut utilizes graph cut statistics as an unsupervised reliability proxy to dynamically gate cross-model supervision, provably guaranteeing a net-positive correction signal across diverse benchmarks.
- Same Words, Different Judgments: How Preferences Vary Across Modalities
-
Addressing the unverified assumption of directly adopting text-oriented protocols for speech alignment, this paper introduces the first rigorously controlled cross-modal preference evaluation benchmark on identical semantic content, revealing that audio judgments feature narrower decision thresholds, reduced length bias, and near-chance cross-modality winner agreement while scaling to good reliability (ICC \(\approx 0.80\)) only with ~9 raters.
- ScaLe-INR: Scale and Learn Implicit Neural Representations
-
To eliminate destructive spectral cross-talk where high-frequency gradient updates corrupt underlying low-frequency manifolds in single-MLP implicit neural representations, ScaLe-INR introduces a multi-branch directional coordinate-scaling architecture guided by Fourier inverse scaling and edge-conditioned sparsity priors, outperforming competitive SOTA baselines across 2D, 3D, and audio tasks.
- Scaling Cheap Synthetic Pretraining Improves Indoor Radio-Map Prediction
-
This paper presents a procedural indoor floor-plan generator and a fast 2D approximate ray tracer to pretrain standard encoder-decoder architectures on 128M synthetic samples, reducing downstream RMSE by an average of 15% across five ICASSP and MLSP benchmark tasks while demonstrating predictable data-scaling power laws.
- scShapeBench: Discovering Geometry from High-Dimensional scRNA-seq Data
-
This paper introduces scShapeBench, the first comprehensive benchmark for automated shape detection in high-dimensional single-cell data (comprising 1,000 synthetic datasets with ground-truth topology and 102 expert-annotated scRNA-seq datasets), and develops scReebTower, a baseline leveraging diffusion geometry to construct Reeb graphs.
- Seahorse: A Unified Benchmarking Framework for Spatiotemporal Event Modeling
-
Introduces Seahorse, a unified benchmarking framework for spatiotemporal point processes (STPPs) featuring an encode–evolve–decode interface, raw-coordinate likelihood reporting with log-Jacobian corrections, and the HawkesNest synthetic stress-test suite to isolate failure modes across heterogeneous model families.
- Seeking the Unfamiliar but Memorable: Conceptual Creativity as Meta-Learning
-
Formalizes conceptual creativity as a Creator-Appraiser bilevel meta-learning game where an adaptive observer's fast inner-loop learning progress acts as an outer-loop reward, guiding a frozen diffusion model to generate visual concepts that are initially unfamiliar yet rapidly learnable.
- Selectivity and Shape in the Design of Forward-Forward Goodness Functions
-
Addressing the long-standing limitation where the Forward-Forward algorithm relies exclusively on sum-of-squares goodness, this paper demonstrates that goodness functions must measure the shape of neural activity rather than total energy, achieving massive gains (up to +72pp over baseline) across six benchmarks via selective and kurtosis-based burstiness functions.
- Self-Consuming Generative Models with Co-Evolving Human Preferences
-
This paper formulates a closed-loop dynamical system capturing the co-evolution of self-consuming generative models and human preferences, proves that pure synthetic retraining causes collapse to singleton equilibria via winner lock-in while sufficient reference data mixing guarantees geometric convergence to a unique globally attracting equilibrium, and develops a lightweight linear programming relaxation to steer long-term diversity at minimal data collection cost.
- Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory
-
This paper rigorously establishes sharp capacity scaling laws for the spectral optimizer Muon on linear associative memory under non-orthogonal Gaussian embeddings and power-law frequencies, proving that one step of Muon recovers \(\widetilde{\Theta}(d^{1+\frac{1}{2\alpha}})\) items—matching second-order Newton's method using only first-order information—while saturating at a substantially larger critical batch size.
- Shift-Aware Identity-Guided Latent Refinement for Referring Audio–Visual Segmentation
-
To address the severe "target shift" problem in referring audio-visual segmentation caused by visual occlusions, acoustic pauses, or distractor interference, this paper proposes an end-to-end framework, SAIL, which introduces a decoupled global semantic anchor and temporal segmentation token design, combined with physical modality evidence residual calibration and a memory-augmented decoder to achieve robust spatio-temporal identity consistency.
- Show Me What You Don’t Know: Efficient Sampling from Invariant Sets for Model Validation
-
This paper introduces a training-free invariance auditing framework that leverages pretrained diffusion/flow-matching models and a feature fiber loss to efficiently sample inputs sharing identical representations, uncovering critical blind spots such as anatomical oversights in vision foundation models.
- SIEVES: Selective Prediction Generalizes through Visual Evidence Scoring
-
Addressing the inability of closed-source frontier MLLMs to expose internal logits and the poor generalization of confidence scores, SIEVES prompts reasoners to expose visual evidence via zoom-in crops in multimodal chain-of-thought reasoning, and trains a lightweight selector to explicitly score localization quality, coherence, and answer correctness, boosting coverage up to three times at low error risk across OOD VQA benchmarks.
- SIGMA: Semantic-Difference Instruction-Grounding Mask Annotator for Text-Driven Image Manipulation Localization
-
Addressing the scarcity of pixel-level annotations for text-driven image manipulation localization (IML), SIGMA combines DINOv2 semantic feature differencing with a structured instruction spatial prior in a dual-branch network, using two-stage noise-calibrated adaptation to automatically extract high-fidelity forgery masks from million-scale unlabeled editing corpora.
- Simple KNN-Based Outlier Detection Achieves Robust Clustering
-
Under the practical assumption that optimal cluster sizes are lower-bounded by a constant factor of the outlier budget, directly discarding points with large KNN distances achieves a constant-factor reduction from robust \(k\)-Means to standard \(k\)-Means without violating center or outlier constraints.
- SimplexUQ: An Evaluation Framework and Benchmark for Conformal Uncertainty on Simplex-Valued Predictions
-
SimplexUQ introduces the first evaluation framework and reproducible benchmark suite (SimplexTasks-12) for measuring coverage allocation failure on simplex-valued predictions, uncovering how global conformal calibration severely under-covers challenging prediction sub-spaces despite nominal marginal validity, and providing a diagnostic-driven wrapper selection heuristic.
- Smooth Partial Lotteries for Stable Randomized Selection
-
Addressing the extreme instability of conventional thresholded partial lotteries in scientific funding and peer review, this paper introduces smoothness as a fundamental design principle and proposes the Clipped Linear Lottery, which achieves bounded Lipschitz shifts under review perturbations and matches the theoretical regret lower bound up to a factor of \((1 - k/n)\).
- Soft Contamination Means Benchmarks Test Shallow Generalization
-
This paper demonstrates that conventional string-matching decontamination fails to filter semantic duplicates ("soft contamination") across massive training corpora, and proves via controlled finetuning on Olmo3 that ecologically valid amounts of soft contamination induce substantial in-benchmark shallow generalization that fails to transfer to independent same-domain benchmarks.
- Soft Geometric Inductive Bias for Object Centric Dynamics
-
This paper introduces object-centric world models equipped with a soft geometric inductive bias using Clifford multivector embeddings, employing non-equivariant Clifford neural networks and block-causal attention to preserve geometric sample efficiency while adapting to symmetry breaking caused by boundaries, collisions, and external forces.
- SOL-ExecBench: Speed-of-Light Benchmarking for Real-World GPU Kernels Against Hardware Limits
-
Addressing the reliance on mutable software baselines and susceptibility to reward hacking in agentic GPU kernel optimization, NVIDIA introduces SOL-ExecBench—a benchmark of 235 real-world kernel problems extracted from 124 frontier models, evaluating performance against analytically derived hardware Speed-of-Light (SOL) physical bounds with an anti-cheat evaluation harness.
- SourceBench: Can AI Answers Reference Quality Web Sources?
-
Introduces SourceBench, the first multi-facet benchmark and calibrated automated evaluator assessing the quality of cited web sources across eight dimensions, evaluating 12 LLMs and search engines over 3,996 sources, and showing that high-quality source retrieval can functionally substitute for complex internal reasoning.
- Sparse All-Layer Connector for Domain Generalised Semantic Segmentation
-
Addressing the bottleneck in multi-foundation-model domain generalised semantic segmentation where conventional depth-aligned fusion overlooks representational hierarchy mismatches and domain sensitivity, this paper proposes the Sparse All-Layer Connector (SALC). By leveraging an adaptive cross-layer selection mechanism based on Sparsegen together with frequency-based candidate dropout regularisation, SALC achieves robust and efficient complementary fusion across unseen target domains.
- Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity
-
Sparse Delta Memory (SDM) scales the hidden state of gated linear RNNs by thousands of times using product-key sparse addressing, matching Gated DeltaNet in parameters and FLOPs while significantly outperforming it in long-context retrieval and matching full attention.
- Sparse Koopman Autoencoders Identify Local Dynamical Regimes in Multibasin Systems
-
Addressing the theoretical limitation that multistable systems generally do not admit a single global continuous Koopman linearization, this paper proposes Sparse Koopman Autoencoders (SKAEs), which use discrete active latent supports as unsupervised local dynamical regime variables to achieve accurate multi-step forecasting and label-free basin identification.
- Sparse Sum-of-Squares Layers for Certified Learning
-
Addressing the computational bottleneck where general differentiable optimization layers destroy forward chordal sparsity during backpropagation, this paper proposes Sparse Sum-of-Squares (SOS) Layers that reuse local clique coordinates via the envelope rule for scalar certified values, achieving up to 2.7x speedup and nearly 100x lower memory usage.
- Spectral Geometry of Attention: From Information Routing to Uncertainty
-
This paper models transformer attention heads as functional mappings between Hilbert spaces of token functions, introduces a Token Difference Operator (TDO) endowed with intrinsic probability geometry to theoretically disentangle attention sinks from representation collapse, and develops SAUQ, an unsupervised single-pass uncertainty estimator that captures upstream routing degradation.
- Spectral Transformer Neural Processes
-
This paper proposes Spectral Transformer Neural Processes (STNPs), which introduce a Spectral Aggregator to dynamically estimate an empirical context spectrum, compress it into a continuous Gaussian mixture, and sample task-adaptive spectral features, injecting an implicit spectral mixture kernel inductive bias into Transformer Neural Processes to effectively model periodicity and quasi-periodicity.
- Speeding up Log-Sum-Exp: Kernel Fusion at the Memory Wall, Integer Arithmetic at the Compute Wall
-
To tackle the redundant high-bandwidth memory round-trips and multi-kernel dispatch overhead in conventional Log-Sum-Exp (LSE) implementations, this paper introduces FuseLSE (a single-pass fused kernel leveraging GPU Special Function Units) and IntLSE (an integer approximation pipeline using shifts and look-up tables), achieving up to 11x speedup on GPUs and 3-6x on CPUs.
- Spherical Flows for Sampling Categorical Data
-
To overcome norm drift and dimension scaling issues in Euclidean continuous flows for discrete sequences, this paper formulates continuous flows on the unit sphere \(\mathbb{S}^{d-1}\) via von Mises-Fisher (vMF) conditional paths, reducing the continuity equation to a 1D scalar ODE via radial symmetry and enabling unified ODE, SDE, and predictor-corrector sampling with a single posterior model.
- SpikeSSL: A Universal Spike Inference Framework with Dynamics-Informed State-Space Layers
-
SpikeSSL incorporates the first-order exponential decay dynamics of calcium transients into bidirectional IIR state-space layers, combining multi-modal conditioning and biophysical simulation augmentation to achieve state-of-the-art zero-shot cross-indicator spike deconvolution with calibrated uncertainty estimation.
- Spin-Weighted Spherical Harmonics Enable Complete and Scalable E(3)-Equivariant Networks
-
To overcome the expressivity loss caused by the angular degree-sum parity constraint in scalar Gaunt Tensor Products (GTP), SpinGTP generalizes tensor operations to Spin-Weighted Spherical Harmonics (SWSH), restoring parity-odd antisymmetric paths while maintaining asymptotic \(\mathcal{O}(L^3)\) computational efficiency and matching full Clebsch-Gordan accuracy.
- SplatLab: Autonomous Scientific Discovery for 3DGS Research
-
SplatLab is the first end-to-end autonomous scientific discovery system demonstrated in the active and technically demanding field of 3D Gaussian Splatting (3DGS), leveraging a verified-reproduction backbone (Splatify/Nerfify), cascade-weighted Elo tournaments on measured GPU PSNR with in-loop falsification, and strict claim-to-log binding to produce manuscripts that beat reproduced targets by +0.18 to +3.78 dB.
- Stability and Diversity of Networked Self-Consuming Generative Ecosystems
-
This paper introduces the first theoretical framework for networked self-consuming generative ecosystems, representing multi-model interactions via directed weighted graphs to establish spectral radius conditions for asymptotic stability, identify dominant cycle gain bottlenecks, and prove graph-Laplacian smoothing bounds on diversity collapse.
- Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping
-
The paper identifies that the high out-of-distribution variance in Looped Transformer length extrapolation stems from spurious correlations between sequence length and loop iterations, and demonstrates that injecting training-time stochasticity via local depth randomization and learned RL-Halting regularizes computation trajectories, sharply shrinking prediction variance and expanding the accuracy–stability frontier.
- Stable and Scalable Probabilistic Numerical Solvers for Stiff and High-Dimensional ODEs
-
Proposes MatfreeEK1, a matrix-free extended Kalman filter ODE solver combining Jacobian-vector products, iterative linear solvers, stochastic covariance estimation, and iterated linearization (IEKF) to achieve linear \(\mathcal{O}(d)\) complexity while preserving full-Jacobian numerical stability on high-dimensional stiff ODEs.
- State-of-art minibatches via novel DPP kernels: discretization, wavelets, and rough objectives
-
This paper introduces novel continuous determinantal point process (DPP) kernels based on multiresolution orthogonal wavelets, achieving an optimal standard error decay rate of \(O(n^{-(1/2 + 1/d)})\) for \(C^1\) objectives while adapting seamlessly to rough functions (Hölder and fractional Sobolev spaces), alongside a general, low-rank continuous-to-discrete conversion pipeline that preserves variance reduction in \(O(Nn^2 + n^3)\) time for discrete minibatches and coresets.
- Statistical Query Lower Bounds for Smoothed Agnostic Learning
-
This paper establishes the first Statistical Query (SQ) lower bound of \(d^{\Omega(1/(\sigma+\epsilon)^2 + \log(1/\epsilon))}\) for smoothed agnostic learning of halfspaces under Gaussian marginals, demonstrating that existing upper bounds via \(L_1\)-polynomial regression are essentially optimal.
- Stealthy World Model Manipulation via Data Poisoning
-
SWAAP is a two-stage data poisoning attack on learned world models that first searches for a harmful yet low-deviation target dynamics model via transition-gradient bilevel optimization, and then realizes it through stealth-constrained gradient matching on fine-tuning transition data, severely degrading planning performance while evading pre-training and robust-training defenses.
- Sticky Jump Diffusions: A Unifying View of Masked, Continuous, and Hybrid Diffusion
-
This paper introduces Sticky Jump Diffusions (SJDs), continuous-time Markov processes whose anchors are token embeddings, mathematically unifying masked, continuous, and hybrid diffusion under flux balance and denoising hazard matching, while opening a novel cross-position blending design space for structured discrete generative modeling.
- Stop the Sampler! Classifier-Based Adaptive Stopping for Sampling Kernels
-
By casting continuous MCMC sampling into a non-acyclic Generative Flow Network (GFlowNet), this paper trains neural state classifiers to adaptively stop trajectories while learning drift-corrected transitions, cutting trajectory lengths by orders of magnitude and substantially improving mode mixing.
- STRABLE: Benchmarking Tabular Machine Learning with Strings
-
STRABLE addresses the severe lack of real-world string-heavy evaluations in tabular machine learning by assembling 108 uncurated datasets across diverse domains, evaluating 445 pipelines to establish that modular pipelines dominate the Pareto frontier while revealing why decoder-only LLM embeddings require variance-normalizing post-processing before dimensionality reduction.
- Strategic Decision-Focused Learning
-
Addressing the core limitation of standard decision-focused learning that assumes an isolated single decision-maker, this paper formalizes strategic decision-focused learning (SDFL), reveals the fundamental non-monotonicity between prediction accuracy and equilibrium payoff, and proposes a hybrid end-to-end training framework combining KKT-directed gradients with strategic checkpoint filtering.
- Strategic Decision Support for AI Agents
-
Addressing the dual vulnerabilities of overconfidence and excessive support-seeking in LLM agents, this paper introduces a reversed decision-support framework and an online oversight algorithm (SOS) that rigorously controls counterfactual missed-support error under distribution-free conditions while dramatically slashing unnecessary support calls.
- Strategic PAC Learnability via Geometric Definability
-
The authors show that even VC dimension 1 classes on the real line can induce infinite VC dimension under simple interval perturbations, and establish that geometric definability over \(\mathbb{R}_{\exp}\) and restricted Pfaffian structures restores strategic PAC learnability with explicit sample complexity bounds.
- StreamPhy: Streaming Inference of High-Dimensional Physical Dynamics via State Space Models
-
To tackle real-time continuous physical field reconstruction under non-uniform, time-varying sparse sensor measurements, this paper introduces StreamPhy, an end-to-end framework coupling an attention-based observation encoder, HiPPO state space recurrence, and a Functional Tensor FiLM (FT-FiLM) decoder provably more expressive than functional Tucker models, achieving 48%+ VRMSE reduction and 20–100× faster inference than diffusion baselines.
- Strong Stochastic Flow Maps
-
Presents Strong Stochastic Flow Maps (SSFMs), an innovative framework that combines a shifted Legendre polynomial expansion of Brownian motion with Chen relations and a semigroup self-distillation objective, achieving few-step simulation-free sampling of strong (pathwise) solutions to additive-noise SDEs.
- Sub-Gaussian Confidence Intervals for Heavy-Tailed Data: Characterizing the Limits of Inference
-
Addressing the impossibility of constructing adaptive confidence intervals for the mean under unknown variance and heavy-tailed asymmetric noise, this paper introduces an "effective sample size" parameter \(\alpha_{n,\delta}\) characterizing truncated variance mass to establish matching upper bounds and non-adaptive minimax lower bounds.
- Support Before Frequency in Discrete Diffusion
-
Proves that under the small-noise limit of discrete diffusion language models, the reverse edit kernel asymptotically decouples into a coarse support-distance scale and a fine within-support frequency coefficient, explaining why models learn valid data support before fine-grained empirical frequencies.
- SurvivalPFN: Amortizing Survival Prediction via In-Context Bayesian Inference
-
Addressing the core challenge of modeling right-censored time-to-event outcomes without specialized architectures and tedious hyperparameter tuning, this paper introduces SurvivalPFN, an in-context Bayesian foundation model pretrained on synthetic identifiable data-generating processes that produces calibrated individual survival distributions in a single forward pass.
- Swift Sampling: Selecting Temporal Surprises via Taylor Series
-
Inspired by predictive coding in the human brain, Swift Sampling models visual latent representations as differentiable trajectories and uses Taylor series to extrapolate expected frame paths, selecting temporal surprises at local maxima with a negligible 1.02× inference overhead while significantly outperforming uniform sampling on long-video benchmarks.
- Swimba: Switch Mamba Model Scales State Space Models
-
Addressing the core computational barrier where naive MoE integration multiplies the recurrent state update cost in selective state space models (SSMs), this paper proposes Switch Mamba (Swimba) to mix expert-dependent parameter streams before recurrence, maintaining a single hidden state trajectory and single-pass evaluation to scale capacity under virtually identical FLOPs.
- SyncLight: Single-Edit Multi-View Relighting
-
SyncLight introduces a pose-free generative framework for consistent multi-view indoor relighting conditioned on a single reference-view edit, leveraging a multi-view diffusion transformer and latent bridge matching to propagate parametric lighting changes across uncalibrated views in a single feedforward pass.
- TabDLM: Free-Form Tabular Data Generation via Joint Numerical–Language Diffusion
-
TabDLM tackles the challenge of generating heterogeneous tabular data containing mixed numerical, categorical, and free-form text features by unifying continuous numerical diffusion and discrete masked language diffusion under a single bidirectional MDLM backbone.
- Tail Competition Explains Scaling in Best-of-N Sampling for Verifiable Problems
-
Reformulates Best-of-N sampling in verifiable problems as an upper-tail competition between correct and incorrect reward distributions, characterizes four scaling phases including variance-driven reward hacking, and introduces an unsupervised tail-tension early stopping rule.
- Tail Cues, Principal Corrections: Plug-and-Play Rectification for Open-Set Test-Time Adaptation
-
Addressing the vulnerability of open-set test-time adaptation (OSTTA) to unknown sample contamination and high computational overhead, TALOR introduces a plug-and-play, backward-free feature rectification module that uses the frozen linear head as a geometric anchor, decomposing features into high-gain principal coordinates and low-gain tail coordinates to estimate and remove affine principal drift online.
- Take It or Leave It: Intent-Controlled Partial Optimal Transport
-
This paper introduces Intent-Controlled Partial Optimal Transport (IC-POT), replacing the classical global scalar rejection rebate with pointwise slack costs on both source and target marginals, thereby allowing statistical bias, predictive confidence, and physical sensor reliability priors to locally price unmatched mass while preserving exact balanced Kantorovich equivalence and dual sparse admissibility.
- Targeted Review for AI-Assisted Biodiversity Surveys: Active Continuous-Score Occupancy Modeling
-
To resolve the vulnerability where imperfect machine learning classifications corrupt downstream population-level ecological inference, this paper introduces Active Continuous-Score Occupancy Modeling (ACORN)—a framework integrating decoupled two-component Gaussian mixture calibration with target expected information gain (Target EIG) to recover oracle-grade occupancy probabilities and covariate relationships with minimal expert verification effort.
- Targeted Synthetic Control Method
-
To overcome finite-sample pre-treatment matching bias in single-treated unit panel data, the Targeted Synthetic Control (TSC) method introduces a TMLE-style one-dimensional exponential weight-tilting update driven by control-unit regression residuals, ensuring exact first-order debiasing while strictly preserving counterfactual boundedness and weight interpretability within the convex hull.
- Task-Induced Riemannian Metrics for Vision Transformer Feature Spaces
-
Addressing the flaw that Euclidean distance and cosine similarity ignore downstream task-sensitive directions in Vision Transformer feature spaces, this paper constructs a task-induced Riemannian pullback metric via decoder Jacobians, introduces a matrix-free spectral diagnostic with a Spectral Pullback Network (SPN), and distills it into a 310K-parameter importance head for high-fidelity token pruning without fine-tuning the ViT backbone.
- Teach-to-Reason: Competition-Guided Reasoning with a Self-Improving Teacher
-
To tackle coarse answer-level rewards and group advantage collapse in reinforcement-learning-based chest X-ray VQA, Teach-to-Reason (T2R) couples Reasoner optimization with an iteratively self-improving Teacher and a case-wise reward mechanism that refines non-degenerate partitions while restoring supervision when group advantages collapse.
- Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care
-
Addresses the scarcity of specialist epilepsy care and Western prescribing bias in low-resource pediatric clinics by introducing Manana, a non-parametric prompt-learning framework with Bayesian Prompt Averaging (BPA) that aligns longitudinal medication regimens and provides calibrated uncertainty-aware deferral to specialists.
- Tempered Guided Diffusion
-
Tempered Guided Diffusion (TGD) formulates training-free conditional diffusion as an annealed Sequential Monte Carlo (SMC) sampler over clean signals, while Accelerated TGD (A-TGD) explores diverse early candidate particles and prunes to a single high-likelihood trajectory partway through sampling, outperforming independent multi-trajectory baselines under matched wall-clock budgets.
- Test-Time Personalization: A Diagnostic Framework and Probabilistic Fix for Scaling Failures
-
To resolve the scaling failures of standard reward models in test-time personalized generation, this paper establishes a predictive unified scaling law governed by four measurable quantities and proposes a probabilistic personalized reward model that scales monotonically and recovers up to 84% of the oracle gap.
- Testable and Actionable Calibration for Full Swap Regret
-
This paper introduces Soft-Binned Calibration Decision Loss (\(\mathsf{SCDL}\)), the first calibration error measure that simultaneously satisfies full actionability against arbitrary swap regret and near-optimal testability with an \(\tilde{O}(T^{-1/2})\) sample estimation rate.
- Testable Learning of General Halfspaces under Massart Noise
-
This paper introduces the first testable learning algorithm for general (non-homogeneous) halfspaces under the Gaussian distribution in the presence of semi-random Massart noise, attaining a sample and computational complexity of \(d^{\tilde{O}(\beta^{-2})\mathrm{polylog}(\min\{1/\gamma, 1/\epsilon\})}\mathrm{poly}(1/\epsilon)\) that qualitatively matches the known Statistical Query lower bound, powered by a novel multiplicative sandwiching polynomial approximation theorem.
- The Bayesian Learning Rule: Beyond KL Geometry
-
This paper generalizes the Bayesian Learning Rule beyond KL geometry by leveraging mixed-space Bregman / Fenchel-Young discrepancies, enabling additive canonical-gradient updates for parameter-dependent bounded-support and heavy-tailed q-Gaussian variational posteriors that scale to modern neural network optimization via q-VON.
- The Causal Description Gap: Information-Theoretic Separations Across Pearl's Hierarchy
-
This paper introduces query-class description length via Kolmogorov complexity to quantify Pearl's causal hierarchy, proving that even with an \(O(1)\)-bit observational description, specifying single-variable interventions or counterfactuals can require \(\Theta(n^2)\) additional bits in the worst case and with high probability.
- The Constitutional Coverage Trilemma in AI Governance
-
This paper uncovers the constitutional coverage trilemma across five core values (safety, helpfulness, honesty, autonomy, and equity), demonstrating that 23 frontier LLM defaults span only ~2% of the human demand hull, and proves that a sparse two-vertex menu cuts mean regret by 47%.
- The Dormant Spiking Neuron: A State-Driven Mechanism for Efficient Spiking Neural Networks
-
Addressing the persistent computational overhead of deeply inhibited neurons in spiking neural networks (SNNs), this paper introduces the Dormant Spiking Neuron (DSN) along with an end-to-end surrogate-gradient BPTT formulation, reducing estimated energy consumption by 40% to over 70% while improving ImageNet baseline accuracy by 1.34%.
- The Future of Facts: Tracing the Factual Generation-Verification Gap
-
This paper systematically traces the factual generation-verification gap (GV-gap) in language models across three training phases (acquisition, continual learning, and updating), showing that verification is learned before generation, is more robust to continual learning, and enters a "multi-verse" state during factual updates where both obsolete and new answers are verified as correct.
- The Limits of AI-Driven Allocation: Optimal Screening under Aleatoric Uncertainty
-
This paper demonstrates that irreducible aleatoric uncertainty imposes an inevitable misallocation bound on purely AI-driven allocation, proves that optimal physical screening must target the margin of algorithmic allocation rather than the highest-risk units, and develops a contractive fixed-point algorithm that substantially improves precision in humanitarian demining and social welfare programs.
- The Reflexivity Threshold: A Phase Transition for Multi-Agent Performative Prediction
-
Introduces the reflexivity matrix \(\Gamma\) with entries \(\Gamma_{ij}=a_{ij}/\mu_i\) for multi-agent performative prediction, establishing a sharp phase transition at \(\rho(\Gamma)=1\) where subcritical systems converge geometrically with bounded cumulative excess loss while supercritical instances force linear cumulative excess loss on any decentralized learner.
- The Score Is Not the Structure: Brain Alignment and Cross-Lingual Transfer
-
The paper introduces a three-step audit framework—content ablation, instrument check, and inference unit correction—revealing that high brain-alignment CKA scores and cross-lingual transfer gradients predominantly stem from metric baseline artifacts and instrument unreliability rather than usable structural correspondence.
- The Score Kalman Filter
-
Addressing the exponential computational explosion of the partition function and the moment closure problem in nonlinear Bayesian moment filtering, this paper introduces the Score Kalman Filter (SKF), which reduces density reconstruction to a single linear solve via score matching and achieves partition-function-free algebraic moment closure and posterior moment recovery via Stein's identity, scaling multimodal continuous nonlinear filtering up to 20 dimensions through pure linear algebra.
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering
-
This paper introduces the Steering Hessian to show that contrastive activation addition (CAA) is fragile because the activation loss landscape is severely anisotropic, and provides a lightweight, norm-preserving Sharp Removal projection that slashes directional curvature by 4–12× and consistently outperforms CAA across 25 open-ended generation conditions.
- The sublevel Flood bifiltration: towards scalable 2-parameter persistent homology
-
Introduces the sublevel Flood bifiltration, which scales 2-parameter persistent homology to large point sets by defining simplex insertion on a compact landmark Delaunay complex covered by full point sets, accelerated via a dual masking heuristic and GPU batched tensor operations.
- Theoretical guidelines for annealed Langevin dynamics in compositional simulation-based inference
-
Addressing the irreducible structural bias induced by running reverse SDEs on aggregated composite scores in compositional simulation-based inference, this paper derives non-asymptotic Wasserstein error bounds for annealed Langevin dynamics under approximate scores, provides explicit closed-form decision rules for step sizes and step counts, and proves theoretically and empirically that Linhart's bridging densities systematically outperform Geffner's in allowable step sizes and computational efficiency.
- Theory-Scale Auto-Formalization of Logics for Computer Science
-
Addressing the gap where existing auto-formalization benchmarks focus solely on isolated statements and lean heavily on pre-existing mathematical libraries, this paper constructs LCS-Bench—the first stand-alone, theory-scale benchmark based on a complete logic textbook comprising 85K lines of Lean 4 code, accompanied by a novel definitional equivalence checker and revealing that frontier LLMs achieve at best 20.1% pass rate on theory-scale formalization.
- TIGER-FG: Text-Guided Implicit Fine-Grained Grounding for E-Commerce Retrieval
-
Addressing the dual modality and granularity disparities in e-commerce image search between cropped queries and full-image item candidates, TIGER-FG introduces an implicit, text-guided fine-grained grounding framework that eliminates explicit object detection, achieving state-of-the-art retrieval with compact 256-dimensional embeddings and 85.7M query parameters through spatial-relational and similarity-distribution distillations.
- To Align or Not To Align: Check Your COMPASS Before You Train
-
Introduces COMPASS, a training-free compatibility assessment score that evaluates local \(k\)-NN neighborhood overlap over a small unannotated probe set, decomposing pair compatibility into an initial "ease" and a paradigm "ceiling" to accurately predict downstream multimodal alignment and model merging performance.
- To discretize continually: Mean shift interacting particle systems for Bayesian inference
-
Targeting continuous posterior densities known only up to an intractable normalizing constant, this paper introduces regularized mean shift interacting particle systems (MSIP) that minimize regularized maximum mean discrepancy (MMD) via scale-invariant preconditioned fixed-point dynamics and inner Gaussian quadrature estimators, delivering robust, mode-collapse-free weighted Bayesian quadrature rules across high dimensions.
- Topological Out-of-Domain Generalization in Dynamical Systems Reconstruction
-
By identifying the mathematical roots of extrapolation failure across bifurcations in hierarchical dynamical systems reconstruction, this paper introduces low-rank sparsity regularization and orthogonal feature splitting alongside a closed-form reliable extrapolation bound, enabling zero-shot forecasting into unseen dynamical regimes across tipping points without fine-tuning.
- Topological Periodicity Test (TopPT) via Confidence Bound of Time-Delay Embeddings
-
Based on time-delay sliding window embeddings and topological data analysis (TDA), this paper rigorously establishes the topological separation (homeomorphic to \(S^1\) vs. contractible interval) and positive reach lower bound between periodic and non-periodic signals, proposing TopPT as the first topological hypothesis test for periodicity with asymptotic Type I and Type II error control.
- Total Variation Distance Estimation in Autoregressive Models
-
This paper formulates and analyzes the problem of estimating total variation (TV) distance between length-\(n\) autoregressive distributions across sample, logit, and noisy logit access models, establishes tight or near-tight query complexity bounds using multilevel Monte Carlo variance reduction, and empirically reveals a 0.586 TV distance between vLLM and SGLang serving identical weights.
- Towards Characterizing Scientific Image Utility and Upgradability
-
This paper introduces the SIU2A framework and curated benchmark to jointly characterize scientific image "utility" (structured error detection and repair feasibility diagnosis) and "upgradability" (consistency-preserving, faithful restoration), uncovering a fundamental failure cascade from visual anomaly perception to closed-loop scientific recovery across leading multimodal foundation models.
- Towards Complementary Keypoint Detection via Mixture of Detectors
-
To overcome the inability of single keypoint detectors to perform across diverse local textures and keypoint budgets, this paper proposes Mixture of Detectors (MoD), which distills heterogeneous detectors into a unified model for single-pass multi-map prediction and deploys a patch-level router conditioned on local texture and global budget to adaptively select optimal detectors, significantly boosting pose and homography estimation accuracy.
- Towards Error-Free EHRs: Reasoning-Intensive Consistency Verification Between Clinical Notes and Structured Tables in Electronic Health Records
-
Addressing the failure of surface-level matching to capture clinical interpretation, composite event relations, and dynamic temporal trends, this paper introduces EHR-ReasonCon—a benchmark with 8,048 expert-annotated entities across 14 MIMIC-III tables—and EHR-Inspector, an LLM-based agent framework that combines note segmentation, two-stage ontology-guided entity extraction, and specialized table-exploration tools, achieving state-of-the-art verification performance under both Harsh and Lenient evaluation criteria.
- TracingFlow: A Simulation-Free Trajectory Inference Framework Based on Second-Order Dynamics
-
Addressing the fundamental inability of first-order flow matching to model crossing trajectories and regulatory momentum in single-cell data, TracingFlow introduces a simulation-free trajectory inference framework grounded in second-order dynamics that formulates dynamical optimal acceleration transport (DOAT) in an augmented state space and completes missing velocities via an exact tridiagonal linear system.
- Training-Free Generative Sampling via Moment-Matched Score Smoothing
-
Proposes Moment-Matched Score-Smoothed Overdamped Langevin Dynamics (MM-SOLD), an interacting particle sampler on a centered and scaled Stiefel manifold that matches target empirical moments at each iteration, eliminating barycentric collapse to achieve fast, training-free, and high-fidelity CPU generative sampling.
- Training a Predictive Coding Network on ImageNet using Equilibrium Propagation
-
This paper bridges Equilibrium Propagation (EP) and Predictive Coding Networks (PCNs) using a modified projected gradient descent (mod-PGD) scheme with asynchronous layer relaxation, successfully training a 10-layer convolutional PCN on full-scale ImageNet to achieve a 13.23% top-5 error rate, closely approaching backpropagation (12.20%).
- Transferable SCF-Acceleration through Solver-Aligned Initialization Learning
-
By differentiating through the self-consistent field (SCF) solver end-to-end to fine-tune initial-guess models (SAIL) and proposing the Effective Relative Iteration Count (ERIC) to account for hidden Fock-build overhead, this work eliminates the supervision misalignment between ground-state fitting and solver dynamics, enabling robust size-extrapolative convergence acceleration across both matrix and coefficient models.
- Transferable SCF-Acceleration through Solver-Aligned Initialization Learning
-
Revealing that the out-of-distribution deceleration of machine-learned matrix initial guesses in Kohn-Sham DFT stems from supervision misalignment with non-linear solver dynamics rather than an extrapolation capacity failure, this paper introduces Solver-Aligned Initialization Learning (SAIL) and the Effective Relative Iteration Count (ERIC) metric, delivering a robust \(1.35\times\) wall-time speedup on large drug-like molecules under hybrid functionals.
- Transformers Linearly Represent Highly Structured World Models
-
By performing an end-to-end mechanistic analysis of an 8-layer Transformer trained on Sudoku solving traces, this paper demonstrates that the model builds an internal linear world model organized around the 27 constraint substructures (rows, columns, boxes) rather than surface grid cells, executing deduction via mid-layer constraint-checking heads and committing via 91 monosemantic naked-single neurons in the final MLP.
- TailOR: Trimming the Long-Tail of Visual World Modeling Evaluation
-
TailOR introduces a multi-dimensional physical interaction benchmark spanning regular, unconventional, and impossible tool-use scenarios to probe whether generative world models truly internalize physical principles, revealing a substantial long-tail performance gap across leading image and video models.
- TailOR: Trimming the Long-Tail of Visual World Modeling Evaluation
-
TailOR introduces a diagnostic benchmark spanning regular, unconventional, and impossible tool-use scenarios under dual predictive and descriptive generation settings, systematically revealing that frontier image and video generative models suffer severe affordance breakdowns and temporal causal inconsistencies due to their reliance on surface statistical co-occurrence.
- TriSearch: Learning to Optimize Triangulations via Bistellar Flips
-
TriSearch formulates high-dimensional polytope triangulation optimization as reinforcement learning over implicit flip graphs, representing moves through circuit-supported local subtriangulations and leveraging simplicial dual-graph convolutions to consistently discover superior triangulations within modest flip budgets and discover significantly more Calabi-Yau threefold topological classes.
- Understanding Circulant Permutation to Extend the CMinHash Estimator
-
By reframing circulant permutations through sampling without replacement and asymptotic expansion for large dimensions, this paper simplifies the complex CMinHash variance derivation into compact closed-form bounds and successfully extends circulant permutations to asymmetric hashing with count-based Minner estimators.
- Understanding Generalization through Decision Pattern Shift
-
The paper introduces Decision Pattern Shift (DPS), a novel perspective that formalizes internal decision mechanisms as channel-wise contribution vectors, theoretically and empirically establishes their strong linear correlation with generalization gap (Pearson \(r > 0.8\)), and constructs a continuous evolutionary spectrum unifying domain shift, OOD, and shortcut learning.
- Understanding Private Evolution as Learning-Augmented Clustering
-
This paper recasts Private Evolution (PE) for differentially private synthetic data generation into a learning-augmented clustering framework, proving that its Wasserstein error bound decouples from the ambient dimension to depend only on the generator's intrinsic doubling dimension, and proposes GAPE (Geometrically Aware Private Evolution) to resolve vote-dilution cluster collapse under provable convergence guarantees.
- Unlocking Compositional Generalization in Continual Few-Shot Learning
-
COMPOSE strictly decouples representation training from compositional inference by leveraging self-supervised ViT patch geometry, training holistically via prototype cross-entropy and cross-correlation decorrelation, and inferring compositionally via bidirectional Chamfer matching over centered slots, achieving 94.14% unseen concept accuracy.
- Unveiling Memorization–Generalization Coexistence: A Case Study on Arithmetic Tasks with Label Noise
-
Using modular arithmetic tasks as a controlled testbed, this work mechanistically dissects the coexistence of memorization and generalization in over-parameterized neural networks under heavy label noise, revealing that networks internally acquire clean algebraic rules while interpolating noise, where frequency filtration recovers near-perfect generalization even under 80% noise while memorization and generalization remain deeply entangled at the neuronal level.
- Unveiling the Value of Motion for Cinematic Camera Trajectories
-
Addressing the geometric entanglement of traditional pose-centric representations with natural language, this paper introduces DirSpeed—a motion-centric parameterization decoupling normalized direction from log-speed—alongside the rich CineScript benchmark and CineGEN, a continuous masked autoregressive generative model that dramatically advances text alignment and physical realism.
- VAANI: Capturing the Language Landscape for an Inclusive Digital India
-
To overcome narrow language coverage, dialectal flattening, and the absence of visual grounding in Indic speech resources, Project VAANI leverages field-captured image prompts to elicit spontaneous speech, creating an open-source multimodal corpus spanning 165 districts, 105 languages, 31,255 hours of audio, and 2,043 hours of expert transcriptions with substantial gains across multilingual ASR benchmarks.
- Validating Causal Abstraction Metrics on Simulated Complex Systems
-
Constructs a standardized benchmark of ten complex systems across discrete/continuous and static/dynamical regimes, evaluating over thirty candidate metrics to show that only causal abstraction error with explicit faithfulness testing reliably separates valid from invalid high-level explanations.
- Variational Trajectory Optimization of Anisotropic Diffusion Schedules
-
This paper introduces a variational framework for anisotropic diffusion models parameterized by a positive semi-definite matrix-valued noise schedule path, deriving an efficient spatial directional-derivative gradient estimator and a second-order matrix Heun solver that consistently improves image synthesis quality over EDM.
- Verifiable Environments Are LEGO Bricks: Recursive Composition for Reasoning Generalization
-
This paper introduces RACES (Recursive Automated Composition for Environment Scaling), which models verifiable environments as domain-codomain compatible "LEGO bricks" that recursively assemble into structurally challenging, programmatically verified tasks under four operators (SEQUENTIAL, PARALLEL, SORT, and SELECT), unlocking strong reasoning generalization across unseen benchmarks and outperforming 300 individual environments with only 50 base environments.
- Vesta: Visual Exploration with Statistical Tool Agents
-
Vesta introduces a framework empowering Vision-Language Model (VLM) agents to dynamically create, persist, and reuse executable statistical diagnostic and visualization tools, substantially boosting probabilistic distribution fitting and time-series modeling performance on the novel Dawn benchmark.
- ViDiC: Video Difference Captioning
-
ViDiC formalizes the task of describing semantic and temporal changes between video pairs, introducing the 1,000-pair ViDiC-1K benchmark with 3,720 checklist items across seven dimensions and an LLM-as-a-judge factual verification protocol, while providing a 60k training set that boosts 7B MLLM accuracy by 11.75%.
- ViT-AdaLA: Adapting Vision Transformers with Linear Attention
-
To alleviate the quadratic complexity bottleneck of vision foundation models (VFMs) at long sequence lengths and avoid the heavy cost of training linear attention from scratch, ViT-AdaLA introduces a progressive three-stage adaptation framework—combining block-level attention alignment and global feature alignment—that allows linearized ViTs to inherit rich visual priors, delivering comparable accuracy to Softmax ViTs while cutting peak memory by 58% and doubling throughput.
- viverra text-to-code with guarantees
-
Viverra introduces a semi-automated formal co-specification framework for text-to-code synthesis that prompts an LLM to co-generate C programs alongside candidate correctness and safety assertions, verifying them compositionally via a portfolio of bounded model checkers to deliver proven behavioral annotations that significantly boost human code comprehension.
- Voronoi-Markov chain and spatial entropy for point pattern analysis
-
Addressing the lack of rigorous null references and multi-scale connectivity metrics in spatial point pattern analysis, this paper introduces a clustering-sensitive dense basin entropy based on Voronoi-Ord distributions and constructs a Voronoi-Markov chain approximating planar Brownian motion, whose spectral gap and entropy rate provide sharp discrimination for spatial point sets.
- Wasserstein Residuals: Learning Gradient Flows from Population Dynamics
-
Addressing the computational bottleneck of optimal transport solvers and poor interpolation across large temporal gaps in conventional JKO schemes, this paper introduces a Wasserstein velocity residual minimization framework alongside a simulation-free stitching algorithm that promotes distribution curves to first-class learnable particle variables, achieving global joint recovery of underlying energy functionals and population trajectories.
- Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery
-
Addressing visual corruption under severe weather, altitude scale discrepancies, and conflicting loss gradients in UAV-satellite cross-view geo-localization, this paper proposes SkyPart—a lightweight swappable head that discovers shared spatial layout via prototype competition and marginalizes altitude nuisances via FiLM, achieving state-of-the-art accuracy with only 26.95M parameters across SUES-200, University-1652, and DenseUAV while demonstrating remarkable resilience on the WeatherPrompt benchmark.
- WebNavigator: Global Web Navigation via Interaction Graph Retrieval
-
Addressing the issue where autonomous web agents suffer from trial-and-error exploration due to lacking global environment topology, WebNavigator constructs dynamic interaction graphs offline at zero token cost and executes a Retrieve-Reason-Teleport workflow online, turning probabilistic exploration into deterministic retrieval and pathfinding to achieve a breakthrough 72.9% success rate on WebArena multi-site tasks.
- Weighted Conformal Clustering
-
Addressing the absence of rigorous point-level uncertainty quantification in traditional clustering and the violation of exchangeability caused by synthetic calibration labels, this paper proposes Weighted Conformal Clustering, which formulates cluster uncertainty as a conditional label-distribution shift problem and introduces a single-fit augmented calibration mechanism to eliminate leave-one-out refitting overhead, constructing statistically valid and compact prediction sets with finite-sample coverage guarantees.
- Welfare, Improvability, and Variance: A Principal-Agent Approach to Optimal Benchmark Item Aggregation
-
Models AI benchmark aggregation as a multitask principal-agent game, proves that optimal item weights follow a closed-form rule \(v^\star=(M+r\Sigma)^{-1}Mw\) governed by social welfare, marginal improvability, and evaluation variance, and demonstrates empirical audits on OLMES items.
- What Cohort INRs Encode and Where to Freeze Them
-
By sweeping freeze depths and applying sparse autoencoder (SAE) dictionary decompositions, this paper reveals that the optimal test-time freeze depth for cohort-trained INRs strictly aligns with the peak weight stable rank layer of the encoder, and mechanistically shows that SIREN learns content-independent spatial tiling primitives while FFMLP memorizes global cohort contours.
- What does a Bayes-filtered transformer believe? A predictive Monte Carlo approach
-
This paper repurposes Predictive Monte Carlo (PMC) as a black-box interpretability tool for Bayes-filtered transformers (BFTs), recovering their internal latent prior and posterior beliefs purely via autoregressive next-token generation and reproducing in-context learning phenomena directly in latent space.
- What Drives Test-Time Adaptation for CLIP? A Controlled Empirical Study from an Update Perspective
-
This paper introduces the TTABC benchmark and organizes 20+ TTA4CLIP methods into three unified paradigms based on what is updated at test time, empirically demonstrating that adaptation gains are predominantly driven by test-time evidence and reliable unsupervised proxies rather than heavy parameter optimization, with no universally optimal paradigm across different distribution shifts.
- What Is Worth Representing? Representational Empowerment for Continual Model Construction
-
Introduces representational empowerment (RepEmp)—formalizing internal model edits as an information channel—and a hierarchical Curator-Actor architecture that enables bounded agents to autonomously construct, retain, and reuse minimal, goal-sufficient representations across changing environments.
- What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion
-
The paper investigates diffusion-friendly latent spaces through the lens of manifold geometry, proposing the Prior-Aligned Autoencoder (PAE) that actively shapes spatial structure, local continuity, and global semantics via targeted prior regularizations, yielding up to 13× faster convergence on ImageNet 256×256 and setting a new state-of-the-art 1.03 gFID.
- When a Zero-Shooter Cheats: Improving Age Estimation via Activation Steering
-
Reveals that vision-language models rely on an "identity shortcut" in zero-shot facial age estimation by retrieving memorized identity knowledge instead of analyzing visual aging features, and proposes an inference-time activation steering method using shallow task vectors to suppress identity retrieval, reducing mean absolute error by up to 25.3% across popular benchmarks.
- When and How to Canonize: a Generalization Perspective
-
This paper develops a covering number framework to establish a rigorous generalization hierarchy among canonization, group averaging, and symmetry-agnostic models, proving that continuous isometric canonizations achieve optimal group-averaged error bounds while discontinuous ones degrade to non-invariant baselines, and providing the first formal proof of polynomial covering number growth for Hilbert curve serialization over exponential lexicographical sorting in point cloud processing.
- When Can Digital Personas Reliably Approximate Human Survey Findings?
-
Evaluating digital personas across 4 prompt architectures, 3 modern LLMs, and 6 reliability dimensions on the Dutch LISS longitudinal panel reveals that synthetic respondents reliably approximate aggregate population distributions but fail to substitute for individual-level prediction or recover multivariate respondent clustering.
- When Copying Is Hard: Copy-Constrained Decoding for Exact Span Reproduction
-
To tackle frequent mid-span token drift and boundary overruns when LLMs reproduce long verbatim spans from context, this paper presents CopyGen, a lightweight plug-in for frozen models that factorizes copying into triggering, adaptive length proposal, and token-level margin-calibrated verification, achieving an average relative exact match improvement of 35% without finetuning alongside over 15x decoding acceleration.
- when do prophets profit in prediction markets
-
Addressing the disconnection where superior probabilistic forecasts fail to generate trading profits in central limit order book prediction markets, this paper proves that proper betting strategies (\(\mathbf{s}_G = \nabla G(\mathbf{p}) - \nabla G(\mathbf{q})\)) derived from strictly proper scoring rules are essentially the unique strategies guaranteeing robust profitability, establishing a tripartite profit decomposition that balances accuracy edges, Bregman bonuses, and liquidity slippage.
- When Does LeJEPA Learn a World Model?
-
This paper theoretically proves that LeJEPA (alignment loss combined with isotropic Gaussian regularization) linearly recovers the ground-truth latent variables from nonlinear observations, and rigorously establishes that the Gaussian is the unique latent distribution that guarantees linear identifiability.
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search
-
This paper reframes the unreliability of rank-1 activation steering in LLMs as an optimization search cost problem, introducing prompt-boundary directional alignment as a geometric prior to prune candidate layers and defining concept granularity to isolate irreducible cross-context rotation from removable noise within the training-free GRACE workflow.
- When Metropolis and Hastings Meet Bradley and Terry: Exact MCMC From Preference Voting
-
Pref-MH introduces the first exact Metropolis-Hastings MCMC sampler that steers generative models using only stochastic binary pairwise preference votes, proving an impossibility result for empirical ratio plug-ins and circumventing it with a provably valid, budget-optimal \(N\)-vote acceptance rule.
- When Stored Evidence Stops Being Usable: Scale-Conditioned Evaluation of Agent Memory
-
This paper introduces a scale-conditioned evaluation protocol for agent memory under evidence-preserving growth, diagnosing memory usability through four trajectory-level metrics: budget-compliant reliability (Pass@B), tail memory-call burden (P90R), failure-regime decomposition, and usable-scale boundary.
- When Symbol Names Should Not Matter: A Logistic Theory of Fresh-Symbol Classification
-
Addressing the fresh-symbol generalization puzzle in template-based classification, this paper develops a logistic theory in the transformer-kernel regime by characterizing accidental token substitutions via colored collision graphs, establishing the first high-probability margin-transfer guarantees.
- Why Do DiT Editors Drift? Plug-and-Play Low Frequency Alignment in VAE Latent Space
-
Investigating multi-turn semantic drift in diffusion transformer (DiT) image editors from a latent frequency perspective, this work reveals that drift stems predominantly from DiT-induced low-frequency deviations rather than VAE round trips, and introduces VAE-LFA—a training-free, plug-and-play low-frequency alignment method that effectively stabilizes multi-turn editing across both white-box and black-box models.
- Why Geometric Continuity Emerges in Deep Neural Networks: Residual Connections and Rotational Symmetry Breaking
-
Reveals the dual mechanisms behind geometric continuity in deep residual networks: residual connections drive cross-layer gradient coherence, while rotational symmetry-breaking non-linearities anchor weights into a shared coordinate frame to prevent drift along \(SO(d)\) orbits.
- Why Latent Actions Fail, and How to Prevent It
-
Addressing the vulnerability of latent action models to exogenous distractors like background motion in uncurated videos, this paper formulates latent action learning under Exogenous Block MDPs, theoretically uncovers two failure modes—future exogenous state leakage and encoder sensitivity amplification—and proves why exogenous-swap reconstruction and robust-target supervision restore action consistency across environments.
- Why Muon Outperforms Adam: A Curvature Perspective
-
This paper takes a first step toward demystifying Muon's training superiority over Adam from a local curvature perspective, proving that while both optimizers share comparable step sizes and first-order gains, Muon achieves significantly smaller normalized directional sharpness (NDS), which substantially mitigates the second-order curvature penalty and balances update energy across curvature modes.
- Workspace-Bench 1.0: Benchmarking AI Agents on Workspace Tasks with Large-Scale File Dependencies
-
To tackle the core challenges of heterogeneous file dependencies and version lineage in realistic office environments, this paper introduces Workspace-Bench, the first large-scale benchmark for Workspace Learning comprising 5 worker profiles, 74 file formats, 388 tasks, and 7,399 rubrics, exposing critical bottlenecks in existing top agents on cross-file and version reasoning.
- World Models as Adversaries: Multi-Agent Self-Play Fine-Tuning for Robust Motion Planning
-
By turning predictive autoregressive world models into role-conditioned adversarial generators, this paper presents Adversarial World Modeling (AWM), a decoupled multi-agent self-play fine-tuning framework that identifies sparse adversarial coalitions via counterfactual credit assignment and optimizes a regret-aware CVaR tail-risk objective with trust-region constraints for robust closed-loop motion planning.
- World Models as Group Actions
-
Formalizes action-conditioned video world models as algebraic group actions on state spaces (e.g., \(SE(2)\) for planar navigation) and enforces identity, inverse, and composition consistency in latent space, eliminating spurious drift and substantially improving multi-step rollout stability without degrading visual fidelity.
- World Motion Models: Flexible Sequence Modeling of SE(3) Trajectories
-
This paper introduces World Motion Models (WMMs), a generative framework that represents dynamic 4D physical scenes as sparse rigid \(\mathrm{SE}(3)\) trajectory sequences and trains a per-token asynchronous flow-matching model with context tokens to seamlessly serve as a policy, forward simulator, and motion planner via marginal conditioning masks.
- WorldCoder-Bench: Benchmarking Physically Grounded 3D World Synthesis
-
Introduces WorldCoder-Bench, the first benchmark for executable and physically grounded browser-native 3D world synthesis comprising 2,026 expert-curated Three.js tasks, and proposes StateProbe, an execution-based verification protocol with mutation-hardened behavioral contracts revealing frontier LLMs' critical shortcomings in dynamic transitions and state synchronization.
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors
-
Reframes video generation evaluation as future world-state prediction conditioned on an initial state and action, introducing the 436-case 22-subcategory WorldReasonBench alongside the ~6K expert-preference WorldRewardBench to expose the persistent gap between superficial visual plausibility and physical/causal reasoning.
- WovenAnchor Matcher: Specialized Intra- and Inter-Image Context Modeling for Feature Matching
-
Addressing the asymmetric demands of intra-image spatial propagation and inter-image correspondence retrieval in semi-dense feature matching, WAM introduces specialized operators—WovenMamba and AnchorCrossAttention—boosting MegaDepth relative pose AUC@5° to 66.1 while reducing parameters and latency.
- X-Palm: Paired Multispectral-to-Smartphone Dataset for Cross-Domain Palmprint Authentication
-
Addressing the prominent cross-domain discrepancy between controlled high-security enrollment and unconstrained mobile authentication, this paper presents X-Palm, the first paired multispectral-to-smartphone palmprint dataset, exposing the performance collapse of existing models under compound real-world variations and establishing a rigorous cross-domain benchmark.
- xHC: Expanded Hyper-Connections
-
Addressing the diminishing returns and cubic compute explosion when scaling Hyper-Connections beyond 4 streams, xHC introduces causal convolution temporal feature augmentation for enriched write-back signals alongside a fixed-plus-routed sparse residual update with dense read, enabling effective scaling to \(N=16\) residual streams and achieving substantial downstream gains on 18B/28B MoE models at modest FLOPs.
- xVGAE: A Hierarchical Variational Graph Autoencoder for Exchangeable Graphs
-
Guided by the Aldous-Hoover representation theorem for exchangeable graphs, xVGAE introduces a hierarchical variational graph autoencoder featuring a global-latent-driven Gaussian mixture node prior and a bilinear graphon decoder, enabling high-fidelity one-shot generation orders of magnitude faster than existing baselines.
- Your Benchmark Is an Empirical Measure Over Difficulty
-
This paper formalizes a benchmark as an empirical measure over item difficulty, theoretically characterizes measure mismatch and local coverage gaps as two primary failure modes, and introduces BenchFill, an evolutionary targeted item rewriting algorithm that fills coverage voids.