💬 LLM (Other)

🔬 ICLR2026 · 56 paper notes

📌 Same area in other venues: 📷 CVPR2026 (2) · 💬 ACL2026 (62) · 🧪 ICML2026 (39) · 🤖 AAAI2026 (29) · 🧠 NeurIPS2025 (53) · 📹 ICCV2025 (6)

🔥 Top topics: LLM ×16 · Diffusion Models ×6 · Reasoning ×3 · Agents ×2 · Layout & Composition ×2

Achieving Olympia-Level Geometry Large Language Model Agent via Complexity Boosting Reinforcement Learning

This paper introduces InternGeometry—the first geometry LLM agent to achieve medal-winning performance. By treating a symbolic engine as a tool, it utilizes an ultra-long-range "Think-Construct/Propose-Verify-Reflect" interaction (over 200 steps per problem) to overcome the lack of heuristics in auxiliary line construction. Combined with Complexity-Boosting RL (CBRL) to progressively increase the difficulty of synthetic problems, it solves 44 out of 50 geometry problems from IMO 2000–2024 using only 13K training samples (0.004% of AlphaGeometry 2), exceeding the average score of gold medalists.

Attend to the Active: Structure-Aware Dynamic Attention in LLMs for Compositional Instruction Following

ATA identifies the structural types of compositional instructions (Chaining/Branching/Paralleling) and decomposes mutually exclusive subtasks in a single forward pass without updating any parameters. By dynamically identifying the currently "active" subtasks at each generation step and masking the "dormant" ones using attention bias, it eliminates interference between subtasks and significantly improves LLM faithfulness to complex compositional instructions.

Best-of-∞: Asymptotic Performance of Test-Time LLM Ensembling

This paper frames majority voting as repeated sampling from a model's answer distribution, investigating the limit accuracy as the number of samples \(N\to\infty\) (termed best-of-∞). It utilizes Bayes factors for adaptive stopping to approximate this limit within finite budgets and formalizes the "optimal weights for ensembling multiple LLMs" as a Mixed-Integer Linear Programming (MILP) problem, demonstrating that ensembling consistently outperforms any single model.

Beyond Magic Words: Sharpness-Aware Prompt Evolving for Robust Large Language Models with TARE

The authors port "Sharpness-Aware Minimization (SAM)" from image/weight space to the discrete textual prompt space, proposing TARE/ATARE: a gradient-free evolutionary framework that "finds the worst paraphrase in the inner layer and selects the most stable neighborhood in the outer layer." This ensures optimized prompts maintain performance under synonymous rewrites, consistently outperforming TextGrad / Revolve across 4 reasoning benchmarks and 5 evaluated models.

Beyond the Known: An Unknown-Aware Large Language Model for Open-Set Text Classification

This paper proposes UnLLM, which reformulates Open-Set Text Classification (OSTC) from "closed-set training + post-hoc OOD detection" into a partition-conditional classification task. By providing LLMs with partial label subsets and explicitly marking samples outside the candidates as "unknown," and employing a three-level "representation-probability-inference" optimization, the model consistently outperforms SOTA in K-F1 / N-F1 across six benchmarks.

BOTS: A Unified Framework for Bayesian Online Task Selection in LLM Reinforcement Finetuning

Proposes BOTS, a unified framework for online task selection in LLM reinforcement finetuning based on Bayesian inference. By fusing explicit evidence (historical pass rates from direct evaluation) and implicit evidence (inferred difficulty of unevaluated tasks via reference model interpolation) with Thompson sampling for exploration-exploitation balance, BOTS achieves up to 50% training acceleration on math, code, and logic tasks with only 0.2% additional overhead.

Breaking the Correlation Plateau: On the Optimization and Capacity Limits of Attention-Based Regressors

This paper provides the first theoretical analysis of the "PCC plateau" phenomenon in attention-based regression models during joint MSE+PCC training. It identifies the root causes as the conflict between MSE optimization and PCC gradients, as well as the capacity upper bound of softmax convex aggregation. It proposes the ECA (Extrapolative Correlation Attention) framework to break these limits using three components: scaled residual aggregation, dispersion-aware temperature softmax, and dispersion-normalized PCC loss.

Cite Pretrain: Retrieval-Free Knowledge Attribution for Large Language Models

By utilizing "Active Indexing" during the continued pretraining phase to bidirectionally bind facts to document identifiers, LLMs can provide verifiable citations while answering in a closed-book setting without any external retrieval, improving citation precision by up to 30.2%.

Compositional-ARC: Assessing Systematic Generalization in Abstract Spatial Reasoning

The authors propose the Compositional-ARC dataset to evaluate the systematic generalization capabilities of models in abstract spatial reasoning—specifically, generalizing from known basic geometric transformations (e.g., translation, rotation) to unseen combinations of these transformations. An MLC-trained encoder-decoder model with only 5.7M parameters reaches 78.26% on systematic tasks, matching the performance of the ARC Prize 2024 winner's 8B model + TTT, while significantly outperforming GPT-4o and o3-mini (<3%).

Constrained Decoding of Diffusion LLMs with Context-Free Grammars

This paper proposes the first constrained decoding method to enforce Context-Free Grammar (CFG) constraints on Diffusion Language Models (DLMs). It abstracts the question of "whether a partial text with holes in any order can be completed into a legal string" as an additive infilling decision problem, reduces it to checking whether the intersection of the target CFG and the regular language of all possible completions is empty, and utilizes highly optimized emptiness checking algorithms to bring the theoretical cubic complexity into a practical range. It increases syntactic correctness to nearly 100% on C++, JSON, and SMILES while slightly improving functional correctness.

COSMOS: A Hybrid Adaptive Optimizer for Efficient Training of Large Language Models

COSMOS decomposes the gradient matrix into a "leading direction + residual" based on feature subspaces. It applies SOAP-style second-order preconditioning to the most informative low-dimensional principal subspace and uses the computationally cheap MUON for the remaining high-dimensional residuals. This achieves pre-training convergence comparable to or slightly better than SOAP while using memory close to MUON (approximately 1/5 of SOAP).

d²Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching

Proposed d²Cache, a training-free approximate KV caching framework for Diffusion-based LLMs (dLLMs). It achieves a \(4.1\times\) inference speedup while improving generation quality through a two-stage strategy: deterministic-prior-guided masked token selection combined with attention-aware non-masked token selection.

Differential Fine-Tuning Large Language Models Towards Better Diverse Reasoning Abilities

This paper discovers that different reasoning abilities (math/code/logic/commonsense) correspond to their own "exclusive" key parameters within LLMs. It proposes DiFT (Differential SFT): first locating key parameter rows for each task using DSR scores, then updating only the union of these key parameters during mix-up fine-tuning, and only the current task's unique key parameters during continual fine-tuning. This approach preserves gains for each reasoning task while avoiding mutual interference during joint training.

Discovering Novel LLM Experts via Task-Capability Coevolution

The AC/DC framework is proposed to facilitate the coevolution of a population of LLMs (evolved through evolutionary model merging) and a set of synthetic tasks (generated by a "Scientist LLM"). This automatically discovers a complementary suite of small expert models in a single run, whose collective Coverage exceeds that of larger models within the same family and even approaches or surpasses GPT-4o, while utilizing significantly fewer total parameters.

Don't Settle Too Early: Self-Reflective Remasking for Diffusion Language Models

Addressing the inherent flaw in masked diffusion language models where "tokens are fixed once unmasked and cannot be corrected," this paper proposes RemeDi. The model simultaneously predicts token distributions and per-token confidence during each generation step. Based on confidence levels, it decides which positions to unmask and which previously generated tokens to revert to masks for resampling. Coupled with a two-stage "Remask SFT + Remask RL" training protocol, it achieves SOTA results among open-source diffusion language models (GSM8K 89.1%, HumanEval 73.2%).

DreamOn: Diffusion Language Models For Code Infilling Beyond Fixed-size Canvas

DreamOn resolves the fixed-length generation constraint of Diffusion Language Models (DLMs) by introducing two special states, [expand] and [delete]. This enables variable-length code infilling without architectural modifications, achieving a 26.4% average improvement on HumanEval-Infilling compared to diffusion baselines and reaching performance parity with state-of-the-art (SOTA) autoregressive models.

Efficient Multi-objective Prompt Optimization via Pure-exploration Bandits

This paper extends prompt selection from single-metric optimization to multi-objective fixed-budget optimization. It proposes two types of algorithms, GENSEC and GENPSI, based on pure-exploration bandits. These methods significantly outperform uniform sampling baselines in summarization tasks and provide theoretical error upper bounds under linear structures.

Eliciting Numerical Predictive Distributions of LLMs Without Auto-Regression

By training lightweight "Magnitude-Factorised Probes" on the hidden states of the final layers of an LLM, the mean, median, and quantiles of the LLM's numerical prediction distribution can be directly recovered without auto-regressive sampling. This approach achieves an inference efficiency equivalent to 20-25 samples with well-calibrated confidence intervals.

ELLMob: Event-Driven Human Mobility Generation with Self-Aligned LLM Framework

The ELLMob framework is proposed based on the Fuzzy-Trace Theory (FTT) from cognitive psychology. It reconciles the competition between user daily patterns and social event constraints by extracting and iteratively aligning "habit gists" and "event gists," achieving explainable event-driven trajectory generation.

Enhancing Persona Following at Decoding Time via Dynamic Importance-Guided Token Estimation for Role-Playing Agents

The Persona Dynamic Decoding (PDD) framework is proposed, which dynamically estimates the context-dependent importance of persona attributes through conditional mutual information and integrates these importance scores into multi-objective reward-guided decoding to achieve training-free inference-time persona following.

Evaluating Text Creativity across Diverse Domains: A Dataset and Large Language Model Evaluator

A context-aware pairwise comparison framework is proposed to evaluate text creativity. The authors constructed the CreataSet dataset, containing 100K+ human-level and 1M+ synthetic data points, and trained the CrEval evaluator, which outperforms GPT-4o by 18.7% in alignment with human judgment.

FACT: Fine-grained Across-variable Convolution for Multivariate Time Series Forecasting

To be added after in-depth reading

Fine-Grained Activation Steering: Steering Less, Achieving More

AUSteer discovers that block-level activation steering is inherently heterogeneous—different dimensions control different token distributions, and mixed steering amplifies both beneficial and harmful signals. It proposes fine-grained steering at the Atomic Unit (AU) level: using activation momentum to locate discriminative dimensions and adaptively adjusting steering strength. By steering \(\leq 100\) dimensions, it significantly outperforms state-of-the-art methods that steer thousands of dimensions.

First is Not Really Better Than Last: Evaluating Layer Choice and Aggregation Strategies in Language Model Data Influence Estimation

This work demonstrates theoretically and experimentally that the conclusion from prior work—claiming the first layer (embedding) is most suitable for influence estimation—is unreliable. It finds that middle attention layers are superior choices and proposes two new cross-layer aggregation strategies, Rank and Vote, along with the Noise Detection Rate (NDR) proxy metric, significantly improving the detection of harmful training samples in LLMs.

How Far Are LLMs from Professional Poker Players? Revisiting Game-Theoretic Reasoning with Agentic Tool Use

This paper systematically analyzes three major reasoning flaws of LLMs in poker (Heuristic Reasoning, Factual Misunderstanding, and the Knowing-Doing Gap) and proposes ToolPoker—the first tool-integrated LLM reasoning system for incomplete information games. By integrating an external CFR solver to provide game-theoretic optimal (GTO) guidance, the framework enables a 7B model to approach Nash equilibrium in Limit Hold'em.

IA2: Alignment with ICL Activations improves Supervised Fine-Tuning

This paper observes that while Supervised Fine-Tuning (SFT) and In-Context Learning (ICL) produce similar outputs, their internal activations differ significantly. Consequently, the authors propose IA2—a self-distillation priming step that uses MSE to align model activations with "ICL-present" activations before performing SFT. This approach significantly improves both few-shot adaptation accuracy and calibration across 12 benchmarks.

In-Context Algebra

This paper designs an in-context algebra task where tokens act as pure variables with meanings randomized for each sequence. It discovers that under this setting, Transformers cease to learn classical Fourier/geometric representations and instead emerge with three symbolic reasoning mechanisms (swap-copy, identity recognition, and closure cancellation). Furthermore, it reveals that these abilities appear sequentially as phase transitions during the training process.

Is the Reversal Curse a Binding Problem? Uncovering Limitations of Transformers from a Basic Generalization Failure

This paper proposes that the "Reversal Curse" is a manifestation of the "binding problem" from cognitive science in Transformers—stemming from inconsistency and entanglement of concept representations. It designs an architecture based on JEPA and memory layers to truly break the reversal curse for the first time without relying on data augmentation.

LLEMA: Evolutionary Search with LLMs for Multi-Objective Materials Discovery

Ours proposes the LLEMA framework, which integrates the scientific knowledge of LLMs with chemical rule-guided evolutionary search and memory-driven iterative optimization. It achieves higher hit rates, stability, and Pareto front quality across 14 multi-objective materials discovery tasks.

Massive Editing for Large Language Models Based on Dynamic Weight Generation

MeG attaches a "dynamic weight neuron" to the LLM, using a diffusion model to generate the neuron's weights conditioned on the knowledge query. This allows massive knowledge editing (1024–10k entries) while adding only a single neuron—simultaneously expanding knowledge capacity and locking interference to the original model as a constant, thereby significantly outperforming existing weight-modification methods on Locality metrics.

Near-Optimal Online Deployment and Routing for Streaming LLMs

This work formalizes the joint problem of streaming online deployment and routing for LLMs for the first time: as new models continuously emerge and old ones become obsolete, the authors propose the StageRoute hierarchical algorithm under a concurrency limit \(M_{\max}\) and cost budget constraints. The study proves a \(\tilde{\mathcal{O}}(T^{2/3})\) regret bound and provides a matching lower bound, achieving near-optimality.

Neologism Learning for Controllability and Self-Verbalization

By adding a "neologism" word embedding to a frozen LLM and training only this embedding to fit examples of a specific concept, one can precisely control model behavior (short, flattery, wrong answers, etc.). Conversely, the model can "self-verbalize" the meaning of this new word in natural language, revealing "machine-only synonyms"—terms that seem irrelevant to humans but stably manipulate machine behavior.

Neural Synchrony Between Socially Interacting Language Models

This study presents the first investigation of neural synchrony between LLMs during social interaction. By training affine transformations to predict the future representations of interaction partners, the authors define the \(SyncR^2\) metric to quantify synchrony strength. Findings indicate that this synchrony depends on social engagement and temporal proximity, and correlates highly with the social performance of LLMs (Pearson \(r\) = 0.88-0.99), mirroring neuroscientific findings of human Inter-Brain Synchrony (IBS).

Optimas: Optimizing Compound AI Systems with Globally Aligned Local Rewards

The Optimas framework is proposed to maintain a Local Reward Function (LRF) aligned with the global reward for each component within a compound AI system. This enables the independent optimization of heterogeneous components (prompts, model parameters, hyperparameters, model selection), achieving an average performance gain of 11.92% across five real-world systems.

Parameters vs. Context: Fine-Grained Control of Knowledge Reliance in Language Models

This paper proposes CK-PLUG, a plug-and-play, training-free decoding-stage method. It uses "Confidence Gain (CG)" to detect conflicts between parametric knowledge and retrieved context at the token level, then blends "parameter-side" and "context-side" probability distributions using a single hyperparameter \(\alpha\). This enables continuous, bidirectional, and controllable adjustment between "the model's own memory" and "retrieved context"—allowing the Memory Recall (MR) on LLaMA3-8B to be tuned anywhere between 9.9% and 71.9% while maintaining generation fluency.

PerFit: Exploring Personalization Shifts in Representation Space of LLMs

PerFit discovers that personalized information in LLMs can be characterized by low-rank representation shifts. By employing a two-stage representation space intervention consisting of "collective shared shifts + personalized exclusive shifts," it approaches or exceeds LoRA/OPPU performance on six LaMP personalization tasks while reducing trainable parameters by approximately 92.3% on average compared to OPPU.

Prompt-MII: Meta-Learning Instruction Induction for LLMs

This paper transforms the task of "writing a high-quality task instruction from a set of examples" into a learnable policy. By meta-training an instruction generator on 3,000+ classification datasets using reinforcement learning, the generator can output a refined instruction for any new task in a single forward pass. The performance matches 100-shot In-Context Learning (ICL) while utilizing 3-13x fewer tokens.

PT2-LLM: Post-Training Ternarization for Large Language Models

Ours proposes PT2-LLM, the first Post-Training Ternarization framework for LLMs. By employing an asymmetric ternary quantizer (comprising Iterative Ternary Fitting and Activation-aware Grid Alignment) alongside a Structural Similarity Reordering strategy, it achieves superior performance at 1.58-bit compared to 2-bit PTQ methods.

ConflictScope: Generative Value Conflicts Reveal LLM Priorities

The authors propose ConflictScope—an automated pipeline for generating and evaluating value conflict scenarios. Given an arbitrary set of values, it automatically generates conflict scenarios between value pairs and evaluates the value priority ranking of LLMs through simulated open-ended user interactions (rather than multiple-choice questions). The study finds that models significantly shift from "protective values" (e.g., harmlessness) toward "personal values" (e.g., user autonomy) in open-ended evaluations, and system prompts can improve alignment with target rankings by 14%.

Rethinking Code Similarity for Automated Algorithm Design with LLMs

This paper proposes BehaveSim, an algorithmic similarity metric based on "Problem Solving Trajectories" (PSTrajs) and Dynamic Time Warping (DTW). It measures algorithmic differences at the execution behavior level rather than the syntax or output level. Integrating BehaveSim into LLM-AAD frameworks like FunSearch/EoH significantly enhances performance.

SIPDO: Closed-Loop Prompt Optimization via Synthetic Data Feedback

SIPDO transforms "data generation" into a real-time feedback signal for prompt optimization: a data generator continuously produces synthetic samples with increasing difficulty that target current prompt weaknesses, while an automatic prompt optimizer diagnoses errors and rewrites prompts per round. This allows prompts to continuously improve and outperform mainstream prompt tuning methods on multiple reasoning benchmarks without relying on external annotations.

Spectral Attention Steering for Prompt Highlighting

Proposes SEKA/AdaSEKA, which learns a "relevance subspace" through spectral decomposition of key embeddings and implements prompt highlighting by directly editing key vectors before attention calculation. This approach requires no storage of the full attention matrix, is fully compatible with FlashAttention, and incurs minimal overhead (+0.03s/sample).

Speculative Actions: A Lossless Framework for Faster AI Agents

Drawing inspiration from CPU speculative execution and LLM speculative decoding, this paper proposes the Speculative Actions framework: while a slow Actor (large model) computes, a fast Speculator (small model) predicts and pre-executes future actions. If the prediction matches, the wait time is skipped to achieve lossless acceleration. The framework reduces latency by 15-30% in scenarios such as Chess, e-commerce, and QA. A confidence-based dynamic branching strategy achieves acceleration effects similar to 3 speculative branches using 40% fewer tokens.

SPRIG: Improving Large Language Model Performance by System Prompt Optimization

SPRIG utilizes an "edit-based genetic algorithm + proxy reward model" to automatically assemble a task-agnostic system prompt. A single system prompt's average improvement across 47 task categories matches "task-specific prompts optimized individually for each task," while the combination of both further achieves new SOTA results and transfers across model families and languages.

Statistical Advantage of Softmax Attention: Insights from Single-Location Regression

By proposing the "Single-Location Regression" (SLR) theoretical framework and employing the order parameter method from statistical physics, this work rigorously proves that softmax attention reaches the Bayes risk at the population level in the high-dimensional limit, whereas linear attention fundamentally cannot. Furthermore, it confirms that softmax consistently outperforms linear attention in finite-sample scenarios, providing the first principled explanation for the superiority of softmax in retrieval tasks.

Stopping Computation for Converged Tokens in Masked Diffusion-LM Decoding

The paper proposes SureLock, which permanently locks token positions in Masked Diffusion LMs once their posterior distributions stabilize (skipping Q projection and FFN, while caching KV). This reduces the per-step attention complexity from \(O(N^2d)\) to \(O(MNd)\), achieving a 30-50% reduction in FLOPs on LLaDA-8B without sacrificing generation quality.

TableMaster: A Recipe to Advance Table Understanding with Language Models

TableMaster systematically decomposes the "structural features" of tables into four categories of challenges. It then provides a "four-flavor recipe"—constructing a Table-of-Focus, verbalizing for semantic enrichment, adaptive switching between textual/symbolic reasoning, and text-guided symbolic reasoning—integrated into a training-free three-stage framework. This approach improves accuracy from 64.73% to 78.13% on WikiTQ using GPT-4o-mini.

Teaching Metric Distance to Discrete Autoregressive Language Models

This paper proposes DIST2Loss, which transforms metric distances between tokens—such as numerical values, coordinates, angles, and VQ codes—into distance-weighted soft target distributions. This allows discrete autoregressive language models to learn structural priors that "near misses are better than far misses" while maintaining the next-token training format, improving data efficiency and downstream performance in visual grounding, robotic manipulation, reward modeling, and image generation.

TEXT2ARCH: A Dataset for Generating Scientific Architecture Diagrams from Natural Language Descriptions

This paper proposes TEXT2ARCH—a large-scale dataset containing 75,000 "architecture image + clean text description + DOT code" triplets. It formalizes the under-explored task of "text description → scientific architecture diagram" as "text → intermediate DOT code → compilation into image." A series of 7B–8B small models are fine-tuned on this data, significantly outperforming DiagramAgent and matching GPT-4o's in-context learning performance.

The Lattice Representation Hypothesis of Large Language Models

Ours proposes the Lattice Representation Hypothesis (LRH) for LLMs: by unifying the Linear Representation Hypothesis (LRH) with Formal Concept Analysis (FCA), it proves that attribute directions in LLM embedding spaces implicitly encode a concept lattice through intersections of half-spaces, thereby bridging continuous geometry with symbolic abstraction.

Toward Safer Diffusion Language Models: Discovery and Mitigation of Priming Vulnerabilities

This paper reveals "priming vulnerability" in Masked Diffusion Language Models (MDLM)—where injecting affirmative tokens during intermediate denoising steps can bypass safety filters. It proposes Recovery Alignment (RA) to train models to recover from contaminated intermediate states to secure responses.

Transducing Language Models

This paper formally elevates the engineering task of "applying deterministic string transformations to language model outputs" to a new class of language models. By encoding transformations using Finite-State Transducers (FSTs) and combining them with pre-trained LMs, the authors propose an algorithm capable of summing over all "source strings mapping to a target string" in finite time. This restores an autoregressive interface (symbol-by-symbol distribution, prefix probability, conditioning) to transformed models without changing any parameters, as demonstrated in "subword-to-byte," "subword-to-word," and "DNA-to-amino acid" scenarios.

Trapped by simplicity: When Transformers fail to learn from noisy features

Research demonstrates that Transformers fail to learn Boolean functions from data containing feature noise—their simplicity bias (tendency to learn low-sensitivity functions) causes models to be trapped by optimal noise predictors that are simpler than the target function, preventing the recovery of the true noise-free target.

VERIFY: A Novel Multi-Domain Dataset Grounding LTL in Contextual Natural Language via Provable Intermediate Logic

VERIFY constructs the first large-scale (200k+ triplets, 13 domains) three-layer aligned dataset of "LTL Formula — Intermediate Technical Language (ITL) — Contextual Natural Language." By employing a pipeline including "enumeration + model checking + provably complete deterministic mapping + LLM generation + multi-level verification," it ensures formal correctness and semantic fidelity. Baselines using T5/Llama/CodeLlama reveal the core challenge: the NL→LTL direction remains extremely difficult (best semantic equivalence is only 31.5%).

WebDevJudge: Evaluating (M)LLMs as Critiques for Web Development Quality

The meta-evaluation benchmark WebDevJudge was constructed to systematically evaluate the capability of (M)LLMs and agentic workflows as judges in web development quality assessment tasks. The study found a ~15% agreement gap between current top-tier models and human experts, revealing two fundamental bottlenecks: failures in identifying functional equivalence and weak feasibility verification.

When Language Models Lose Their Mind: The Consequences of Brain Misalignment

This paper uses "brain misalignment" interventions to deliberately remove information from LLM representations that predicts fMRI signals in the human brain's language areas while maintaining language modeling loss. It finds that such a decrease in brain alignment systematically impairs over 200 language probing tasks, including semantics and syntax, whereas enhancing brain alignment yields linguistic performance gains.