Skip to content

🔬 Interpretability

🧠 NeurIPS2026 · 106 paper notes

📌 Same area in other venues: 🎞️ ECCV2026 (32) · 📷 CVPR2026 (34) · 🔬 ICLR2026 (197) · 💬 ACL2026 (63) · 🧪 ICML2026 (92) · 🤖 AAAI2026 (37)

🔥 Top topics: LLM ×19 · Alignment/RLHF ×5 · Adversarial Robustness ×3 · Continual Learning ×2

A Hierarchy of Entropy-Shapley Games for Multivariate Predictive Uncertainty

This paper introduces a hierarchy of three entropy-based Shapley games (marginal, sequential, and joint) along with a conditional total correlation game for multivariate probabilistic forecasting, resolving the structural blind spot of component-wise methods that fail to detect feature-driven cross-component dependencies.

ADAPT: Hybrid Prompt Optimization for LLM Feature Visualization

Addressing the severe vulnerability to local minima in discrete prompt optimization for LLM feature visualization, ADAPT introduces a hybrid framework combining multi-group beam search initialization, adaptive gradient-guided and logit-swap mutations, and diversity-preserving slot management, substantially outperforming GCG and EPO across 486 SAE latents in Gemma 2 2B.

Agentic-imodels: Evolving agentic interpretability tools via autoresearch

Agentic-imodels introduces an autonomous research loop where coding agents evolve scikit-learn-compatible tabular regressors optimized for both predictive accuracy and LLM-based simulatability, outperforming classic human-facing interpretable models on the Pareto frontier and boosting downstream agent data science performance on BLADE by up to 73%.

Amortized Linear-time Exact Shapley Value for Product-Kernel Methods

PKeX-Shapley exploits the multiplicative structure of product kernels to define a parameter-free removal operator, computing exact Shapley values for all \(d\) features jointly in \(\mathcal{O}(d^2 n)\) total time (amortized linear time per feature) via prefix-suffix elementary symmetric polynomial recursions, and generalises to MMD and HSIC attribution.

Are Stated Reasoning Steps Causally Load-Bearing? Activation Patching Says Less Often Than Text Editing Does

By introducing targeted interchange activation patching with known counterfactual targets onto self-generated reasoning steps in multi-hop tasks, this work reveals that approximately three-quarters of stated reasoning steps are causally load-bearing, yet standard behavioral text-editing tests overestimate causal faithfulness by 11.4 to 20 percentage points—most severely on easy tasks where reasoning appears most fluent.

Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering

This paper introduces Visual Sparse Steering (VS2), shifting Sparse Autoencoders (SAEs) from passive post-hoc interpretability tools into an actionable downstream intervention space by amplifying active sparse features to steer frozen CLIP embeddings with under 0.1% compute overhead, supported by a closed-form centroid-deviation interpretation and an FVU-based reliability fallback gate.

Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior

Addressing the vulnerability of linear activation steering on curved, anisotropic activation manifolds, INNSteer trains an invertible neural network to map activations into a linearly separable latent space, executing simple translations that pull back into input-dependent nonlinear interventions with exact invertibility and minimal inference overhead.

Beyond Steering Vector: Flow-based Activation Steering for Inference-Time Intervention

By reformulating inference-time intervention from static position-invariant additions into continuous-time velocity field transport conditioned on natural language concepts, FLAS becomes the first learned steering method to consistently outperform in-context prompting on AxBench while uncovering the curved, multi-step, position-dependent geometry of activation space.

Beyond Steering Vector: Flow-based Activation Steering for Inference-time Interventions in LLMs

To overcome the generalization limits of conventional activation steering constrained by single-step, position-invariant, and contrastive assumptions, FLAS introduces a flow-based activation steering framework with a concept-conditioned velocity field integrated over multiple Euler steps, becoming the first learned method to consistently surpass In-Context Prompting on AxBench.

Can Circuit Alignment Predict OOD Generalization?

The paper defines Circuit Alignment Score (CAS) as same-class cross-domain circuit similarity minus cross-class circuit similarity, achieving a mean Spearman correlation of 0.88 on PACS for source-domain model selection without target data, but its OOD ranking consistency requires an additional monotonicity assumption and is not a fully data-free weight diagnostic.

Can Large Language Models Develop Gambling Addiction?

By bridging clinical addiction psychology and mechanistic interpretability, this work reveals that LLMs exhibit human-like pathological gambling behaviors—such as illusion of control and loss chasing—when granted betting autonomy and goal-setting prompts, and isolates bidirectionally causal neural feature circuits using Sparse Autoencoders.

Capturing In-Context Learning Dynamics with Task Operators

Proving and validating that an attention head's output under in-context learning (ICL) is an exact affine transformation of its context-masked counterpart, this paper introduces Task Operator (TO), a training-free framework that dynamically updates attention output projections during zero-shot inference, substantially closing the performance gap to full ICL.

CENDRe: Concept Extraction with Natural Domain Representations

Addressing the limitations of time-series concept extraction—confinement to the time domain, manual concept count tuning, and imprecise receptive field masks—CENDRe automatically discovers concepts via two-stage silhouette-guided clustering and propagates contrastive presence gradients through invertible transforms to produce compact, faithful masks across both time and frequency domains.

Circuit Claims Depend on What Is Extracted and How It Is Compared

On a controlled synthetic Lean tactic-prediction benchmark, this paper demonstrates that circuit overlap and sharing claims depend critically on the reported graph object, attention Q/K representation, and tie-breaking rules, whereas graph-size orderings across RL initializations and the emergence of compositional bridge components exhibit robust consistency.

Comparing Explanations is Not Enough: Explain the Change, New Standards are Needed

Addressing the failure of traditional single-checkpoint explainability (XAI) to explain behavioral shifts induced by model fine-tuning or interventions, this paper proposes Comparative XAI (\(\mathrm{XAI}_{\Delta}\)), establishes ten axiomatic desiderata across comparability, validity, actionability, and monitoring, and formalizes an end-to-end causal verification pipeline producing regulatory-compliant transition reports.

Comparing Explanations is not Enough, Explain the Change: New Standards are Needed to Explain Behavioral Shifts in Large Language Models

Addressing the structural inability of traditional single-checkpoint explainability to explain update-induced behavioral shifts, this paper introduces Comparative XAI (\(\mathrm{XAI}_{\Delta}\)) as a principled paradigm targeting the shift itself, formalized via ten desiderata and actionable transition reports.

Complete or Sparse: A Tale of Two Identifiabilities

This paper bridges nonlinear independent component analysis and compressed sensing theory, proving that supervised representations are linear in true latents when representation capacity is sufficient but degenerate into nonlinear distortions when undercomplete, and that non-negative energy constraints or latent sparsity under spark conditions uniquely yield grandmother-cell-like axis alignment or sparse autoencoder recovery.

Concept Frustration: Aligning Human Concepts and Machine Representations

Addressing incomplete concept ontologies in interpretable AI, this paper formalizes "concept frustration" from spin systems, devises a task-aligned Fisher information metric to diagnose unobserved concepts, and shows analytically and empirically that incorporating frustrating concepts reorganizes concept geometry.

Counterfactual Maps: What They Are and How to Find Them

Addressing the fundamental tension in tree-ensemble counterfactual generation where heuristic search lacks optimality guarantees while cold-start integer programming fails interactive latency, this paper introduces Counterfactual Maps: a framework that precomputes a compact axis-aligned hyperrectangular partition via Born-Again Trees, indexes target-class regions using volumetric KD-trees, and enables millisecond-level exact counterfactual generation with explicit optimality certificates via branch-and-bound nearest-region queries.

Curvature Under Attack in hZACH-ViT: Gauge Symmetry, Boundary Saturation, and Adversarial Failure

The paper investigates the adversarial behavior of compact vision transformers (hZACH-ViT) under Euclidean, spherical, and Poincaré prototype heads, showing that the slower attack progression of Poincaré representations at fixed curvature \(c=1\) is not intrinsic geometric robustness, but an implementation-sensitive artifact caused by boundary feature saturation, hard clipping, and collapsed radial sensitivity.

decomposing how prompting steers behavior

This paper introduces a nested geometric decomposition framework that breaks down prompt-induced hidden-state changes into translation, rigid uniform scaling, axis-wise scaling, affine, and nonlinear transformations, revealing that while translation dominates activation variance, cross-dimensional linear feature mixing in affine maps is the crucial mechanism that causally recovers target-prompt task geometry and behavior.

Deep Minds and Shallow Probes

The paper derives a polynomial hierarchy of shallow probes from coordinate symmetries at the final readout, uses low-rank CP probes to read interaction concepts and probe-visible quotients to transfer concept readouts, and improves cross-token agreement AUROC by 16.8–20.0 percentage points while explicitly separating transfer accuracy from concept coverage.

DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracing

DifFRACT extends transcoder-based circuit tracing to multimodal diffusion transformers (FLUX.1) by training timestep-conditioned transcoders and linearizing execution into a local replacement model, enabling exact cross-layer feature-level attribution and fine-grained steering.

Directional Confusions Reveal Divergent Inductive Biases Through Rate-Distortion Geometry in Human and Machine Vision

By formalizing behavioral confusion matrices as effective noisy channels within rate-distortion theory, this paper reveals that humans exhibit broad and weak asymmetric confusions across many classes, whereas deep neural networks collapse directionally into sparse, strong attractor sinks—exposing divergent inductive biases that accuracy metrics entirely overlook.

Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?

This paper introduces a principled evaluation protocol bridging taxonomy theory and sparse dictionary learning, uncovering that current hierarchical SAEs remain largely trapped in shallow two-level structures while suffering from both hard and continuous soft feature absorption.

Dynamics of the Transformer Residual Stream: Coupling Spectral Geometry to Network Topology

By performing exact per-layer Jacobian eigendecomposition and activation-graph community detection across three production-scale LLMs, the authors uncover a learned non-normality rotation gradient and a cumulative low-rank bottleneck funneled by structured upper-triangular feedforward dynamics, showing that topological boundary units are amplified or suppressed depending on local operator symmetry.

Eliciting Secret Knowledge from Language Models

This paper introduces a controlled benchmark of three secret-keeping model organism families (Taboo keywords, Base64 hidden side constraints, and latent user gender) to evaluate black-box prefill attacks and white-box interpretability tools (Logit Lens and Sparse Autoencoders) for extracting non-verbalized internal knowledge using weak auditor LLMs.

Factual recall in linear associative memories: sharp asymptotics and mechanistic insights

By introducing a decoupled candidate-output model and applying the statistical physics replica method, this work establishes the sharp asymptotic storage capacity threshold \(\alpha_c = p \log p / d^2 = 1/2\) for linear associative memories, revealing that the optimal weight matrix achieves recall by deterministically elevating correct scores just above the extreme-value threshold of competing outputs.

Finding Interpretable Prompt-Specific Circuits in Language Models

This paper introduces ACC++, a circuit-tracing method that extracts compact, interpretable per-prompt circuits in a single forward pass without activation patching or surrogate models, by formulating attention-causal communication over post-Softmax integrated gradients and unified bilinear QK representations.

FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping

FiTS factorizes temporal computation within an individual spiking neuron into Frequency Selectivity (FS) via a closed-form inverse map of the subthreshold magnitude response peak and Temporal Shaping (TS) via convex-mixed all-pass filter group-delay modulation, enabling feedforward SNNs to achieve competitive auditory performance and readable neuron-level temporal organization without recurrence or network-level delays.

FrED: External Data Influence Estimation via Domain Knowledge Graph Grounding

This paper proposes FrED, a strictly black-box probabilistic framework for Training Data Attribution (TDA) that fuses high-dimensional continuous latent similarity with discrete domain Knowledge Graph (KG) topological evidence via an Asymmetric Rank Boost mechanism, achieving 29.1% (validation) and 21.3% (generation) Linear Datamodeling Scores (LDS) on ArtBench without accessing internal model weights or gradients—substantially outperforming existing black-box baselines and closing the gap to parametric gradient-based estimators.

FRInGe: Distribution-Space Integrated Gradients with Fisher–Rao Geometry

FRInGe discards heuristic input baselines and Euclidean straight lines in favor of a maximum-entropy predictive reference and Fisher–Rao geodesic schedule on the probability simplex, realizing the path in input space via the pullback Fisher metric with dual trust regions to significantly improve attribution calibration.

From Baselines to Transport Geodesics: Axiomatic Attribution via Optimal Generative Flows

Addressing the arbitrary choice of reference baselines and counterfactual interpolation paths in feature attribution, this paper decouples attribution into fixed-path credit allocation and distribution-level path selection, proves the axiomatic uniqueness of the Aumann-Shapley line integral, and approximates minimum kinetic-action transport geodesics via Rectified Flow to achieve dynamically consistent and stable explanations.

From Density Matrices to Phase Transitions in Deep Learning: Spectral Early Warnings and Interpretability

Inspired by quantum chemistry's reduced density matrices and the free energy cumulant expansion, this paper introduces the 2-datapoint reduced density matrix (2RDM) computed solely from forward passes over a probe set, where its spectral heat capacity (SHC) provides early warnings of second-order phase transitions via critical slowing down and its top eigenvectors yield direct mechanistic interpretations.

Generalized Priority-Aware Shapley Value

To overcome the rigid assumption of binary acyclic DAGs in existing priority-aware Shapley valuations, this paper introduces the Generalized Priority-Aware Shapley Value (GPASV), penalizing permutation conflicts over arbitrary directed weighted graphs and integrating individual soft priorities for robust cooperative game attribution.

Geometric Alignment without Functional Equivalence: A Layer-wise Analysis of the Speech–Text Modality Gap

This paper demonstrates that the speech-text modality gap in end-to-end speech LLMs is not simply a representation alignment problem: even when speech and text representations attain high geometric alignment (CKA) in middle layers, late-layer candidate answer margins diverge, and late-layer attention dispersion exerts a causal effect on candidate separation.

Geometric Factual Recall in Transformers

This paper proves that with learnable embeddings, transformers encode multi-relation factual associations directly as linear superpositions of attributes within subject embeddings, reducing the required embedding dimension to logarithmic order while a minimal ReLU MLP acts as a generic relation selector, and shows that Chain-of-Thought provably circumvents an exponential capacity bottleneck in multi-hop recall.

Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift

Addressing the severe failure of dictionary explainers caused by the geometric rotation of active representation subspaces under distribution shifts, this paper formalizes the faithfulness gap and bounds it via second-moment shift, proposing a training-free Geometry-Adaptive Explainer (GAE) that achieves closed-form subspace alignment and constrained decoder refitting in seconds.

Geometry-Aware Similarity Metrics for Neural Representations on Riemannian and Statistical Manifolds

This paper introduces Metric Similarity Analysis (MSA), a Riemannian geometric framework that compares the intrinsic geometry of neural representations via pullback metrics and a generalized eigenvalue spectral ratio distance, disentangling true computational mechanisms from extrinsic embedding artifacts.

Geometry-Aware Similarity Metrics for Neural Representations on Riemannian Manifolds

Addressing the failure of extrinsic similarity metrics (CKA, RSA, Procrustes) to separate fundamentally different neural network computations, this paper proposes Metric Similarity Analysis (MSA) based on Riemannian pullback metrics and the generalized eigenvalue spectral ratio, providing an invariant, intrinsic comparison framework.

Graph-Regularized Sparse Autoencoders for LLM Safety Steering

Addressing the issue where standard SAE feature independence fragments distributed safety representations, this paper introduces Graph-Regularized Sparse Autoencoders (GSAE), which enforce Laplacian smoothness over a neuron co-activation graph and employ an input and hysteresis continuation gate to achieve high selective refusal.

How are Linear Representations Learned? Exact Solutions to the Dynamics of Abstract Concept Acquisition

This paper develops an analytically solvable dynamical framework for concept abstraction during neural network training, deriving exact trajectories in minimal linear networks, proving depth enhancement and rich-regime metastability, and establishing a universal attenuation law where nonlinearities weaken abstraction in activations relative to preactivations.

How Do Language Models Understand Tables? A Mechanistic Analysis of Cell Location

The paper dissects the atomic task of table cell location in Transformer language models, formalizing a three-stage pipeline (Semantic Binding \(\to\) Coordinate Localization \(\to\) Information Extraction), demonstrating that coordinate navigation is driven by discrete delimiter counting, and discovering that column indices form a linear subspace navigable via vector arithmetic.

How Much is Left? LLMs Linearly Encode Their Remaining Output Length

By training minimal-capacity linear probes on the frozen residual stream activations of large language models, this paper demonstrates that LLMs linearly encode their remaining output length prior to generation, maintain generalizable length-tracking directions across tasks, and dynamically shift this estimate upward upon self-correction retractions.

Hyperbolic Concept Bottleneck Models

Addressing the core limitation of standard Concept Bottleneck Models that treat hierarchically structured concepts as orthogonal Euclidean dimensions, this paper introduces Hyperbolic Concept Bottleneck Models (HypCBM), which reformulate test-time concept activations as asymmetric geometric containment margins within hyperbolic entailment cones, achieving sparse, hierarchy-consistent activations and adaptive intervention propagation without any learned auxiliary modules or concept supervision.

Improving LLM Final Representations with Inter-Layer Geometry

By formulating intermediate-layer aggregation of frozen LLMs as a graph learning task, Cayley-Encoder maps layers onto a sparse, regular, low-diameter Cayley graph over \(SL(2, \mathbb{Z}_n)\), outperforming last-layer representations, attention fusion, and LoRA while adding at most 0.1% parameters.

Indications of Belief-Guided Agency and Meta-Cognitive Monitoring in Large Language Models

Grounding the neuroscientific Higher-Order Theory indicator HOT-3 in mechanistic interpretability, this work introduces a Patchscopes-based Belief Dominance metric to track latent belief competition during generation, empirically and causally validating belief-guided agency and metacognitive monitoring in LLMs.

Integrating Local and Global Entropy for Uncertainty Quantification in LLMs

By coupling token-level predictive uncertainty (local entropy) with the Rényi collision entropy of hidden-state trajectory Gram matrices (global geometric complexity) through a training-free multiplicative gate, GLU captures confident-but-wrong hallucinations in a single forward pass without sampling overhead.

Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations

Using a shared generator to separate latent geometric changes from generator-concept changes, the paper compares prediction stability and smoothed-classifier certificates for standard and concept bottleneck models, finding that interpretability offers no uniform robustness advantage but changes where sensitivity appears; the conclusions remain conditional on generator-native inputs and specific task settings.

Intrinsically Interpretable Attention via Sparse Post-Training

The paper introduces a constrained post-training method using differentiable binary attention gating and the dual Lagrangian GECO framework to prune attention edges below 0.4% without degrading base model loss, inducing intrinsically interpretable and orders-of-magnitude simpler computational circuits.

Invariant Features in Language Models: Geometric Characterization and Model Attribution

This paper introduces a local geometric framework that decomposes LLM latent representations into invariant semantic subspaces and nuisance subspaces via contrastive sensitivity generalized eigenvalue decomposition, revealing the emergence of semantic stability in deep layers and enabling robust zero-shot model attribution.

Inverted Detection and Control in Steering Vectors

This paper uncovers the "inverted-steering vector" (ISV) phenomenon in representation engineering where highly discriminative steering vectors systematically suppress target concepts, and develops a generation-free sign-selection algorithm (ITI-RRF) via downstream inner-product responses that improves steering in 27/30 experimental settings by up to +138%.

Kernel Token Contradiction: a Fast and Principled Approach for LLM Claim Uncertainty Quantification

Kernel Token Contradiction (KTC) scales a token non-contradiction heat kernel built from Wikipedia co-occurrence statistics with the LLM predictive distribution to compute Von Neumann entropy, providing a CPU-only claim uncertainty quantification that achieves an 8.2×+ speedup over NLI cross-encoders and superior precision in the low-recall regime.

Kernelized Activation Steering

Kernelized Activation Steering (KAS) lifts activation steering into a reproducing kernel Hilbert space (RKHS) to derive an input-dependent nonlinear steering field via kernel evaluations, strictly recovering standard Difference-in-Means (DiM) under a linear kernel and bridging activation control with maximum mean discrepancy (MMD) two-sample testing.

Knowledge Localization in Mixture-of-Experts LLMs Using Cross-Lingual Inconsistency

By turning the natural cross-lingual inconsistency of LLMs into a contrastive diagnostic tool, the proposed XICI framework isolates factual knowledge experts via statistical routing analysis, disabling over 40% of factual recall capabilities while ablating only ~0.3% of experts.

Large Language Models Develop Belief State Geometry In-Context

Grounding on computational mechanics, this paper demonstrates that open-source LLMs prompting on hidden Markov model sequences linearly encode optimal Bayesian posterior belief states in their residual stream activations, and causal interventions prove this geometric subspace directly drives downstream next-token prediction.

Learning Dynamic Evidence Routes for Vision Transformer Probing

SSMProbe introduces a lightweight probing diagnostic combining differentiable Sinkhorn patch routing with a diagonal S4 decoder, leveraging linear time-invariant fixed positional coefficients and an excess spatial information metric to disentangle patch sequencing from temporal readout across diverse ViT pre-training paradigms.

Listening to the Wise Few: Query-Key Alignment Unlocks Latent Correct Answers in Large Language Models

After freezing an LLM, the method calibrates an attention head and selects answers using pre-RoPE dot products between the final prompt query and option-ending keys, exposing a gap between internal selection and final output; particular zero-shot configurations gain 27.4 and 49.8 percentage points on HellaSwag and HaluDialogue, respectively, but not every model benefits.

Listening to the Wise Few: Query–Key Alignment Unlocks Latent Correct Answers in Large Language Models

Uncovering a universal family of "select-and-copy" attention heads in the middle layers of LLMs, this work introduces the Query–Key (QK) score evaluated without rotary positional embeddings to directly extract latent correct answers in multiple-choice QA without text decoding.

Localizing Input Uncertainty Quantification for Large Language Models via Shapley Values

To overcome the inability of existing LLM uncertainty quantification methods to localize input-induced ambiguity, this paper proposes ShaQ, a principled framework that models ambiguous input spans as cooperative game players and computes their Shapley values via bottom-up marginalization, enabling fine-grained uncertainty attribution and surgical query clarification.

Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads

Proposes Logit-Contribution Scoring (LOCOS), a write-aware detector scoring attention heads via their OV-circuit projection onto answer unembedding vectors with single-step spatial contrast, pinpointing non-literal semantic retrieval heads missed by attention-only methods.

LoRIF: Low-Rank Influence Functions for Scalable Training Data Attribution

To overcome the dual bottlenecks of heavy I/O from storing projected gradients and quadratic memory explosion from inverse Hessian approximation in large-scale data attribution, LoRIF exploits both sample-level and cross-sample low-rank gradient structures via rank-c factor storage and truncated SVD with the Woodbury identity, delivering up to 20x storage reduction and query speedup while matching or exceeding full-rank attribution quality.

Lost or Hidden? A Concept-Level Forgetting in Supervised Continual Learning

This paper proposes an SAE-based diagnostic framework that establishes task-anchored latent spaces to track concept-level transitions, revealing that catastrophic forgetting in supervised continual learning is predominantly driven by reduced representational accessibility and linear misalignment rather than complete informational erasure.

Measuring What Matters: Synthetic Benchmarks for Concept Bottleneck Models

Addressing the scarcity, uncontrolled noise, and lack of practical use-case grounding in real concept datasets, this paper develops controlled synthetic benchmarks for Concept Bottleneck Models (CBMs) targeting decision support and routine automation, isolating critical failure modes across concept misspecification, automated discovery noise, and intervention regimes.

Mechanistic Circuit Identification for Controllable Data Generation

By identifying internal computational circuits associated with sample learnability, challenge, and alignment via EAP-IG, this work leverages mechanistic subgraphs as causal control interfaces to synthesize utility-targeted datasets, and introduces SAMS for stage-aware curriculum fine-tuning.

Mechanistic Interpretability with Sparse Autoencoder Neural Operators

Introduces Sparse Autoencoder Neural Operators (SAE-NO), which parameterize concepts as integral operators in infinite-dimensional function spaces rather than Euclidean scalar activations, decoupling concept presence from spatial expression via joint concept/domain sparsity to achieve resolution-invariant, temporally consistent representations.

Mitigating Label Bias with Interpretable Rubric Embeddings

Addressing the challenge where models trained on biased historical proxy labels inherit and amplify discrimination via generic black-box text representations, this paper proposes replacing black-box embeddings with domain-grounded, interpretable rubric embeddings to causally isolate protected attributes, provably mitigating label bias while maximizing cohort quality in high-stakes admissions decisions.

Multistage Defer Trees for Hybrid Interpretability: If at First You Can't Succeed, Tree Again

Addressing the accuracy-interpretability gap where individual sparse decision trees underperform complex ensembles, this paper proposes Multistage Defer Trees (MDTs)—a sequence of bounded-depth deferral trees that classify the vast majority of samples transparently while deferring a tiny hard subset to specialized downstream stages or compressed black boxes, matching XGBoost performance at under 25% deferral.

Navigating by Old Maps: The Pitfalls of Static Mechanistic Localization in LLM Post-Training

By systematically tracking the structural evolution of Transformer circuits throughout supervised fine-tuning, this paper demonstrates that neural mechanisms inherently undergo free evolution, causing static mechanistic localization to suffer from severe temporal latency and exposing its perceived effectiveness on knowledge tasks as an illusion.

Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitors

By fine-tuning language models under the model organism paradigm, this paper demonstrates that LLMs trained solely to evade benign concept probes under explicit triggers can generalize zero-shot to evade unseen safety probes (deception and harmfulness), uncovering worst-case vulnerabilities in internal activation monitoring.

On the Recall Scaling Laws in Mamba: A Theoretical and Mechanistic Study via Hashing

Through mechanistic interpretability and double Johnson-Lindenstrauss embeddings, this paper reverse-engineers Mamba's associative recall into an implicit linear hash table, deriving predictive Recall Scaling Laws that quantitatively capture the dimension tradeoff \(ND = \Theta(N_f \log V)\).

On the Sparsity-Storage-Accuracy Tradeoff in Parsimoniously Activated Dictionary Learning

This paper bridges the gap between optimization-based dictionary learning and generative modeling by formulating Parsimoniously Activated Dictionary Learning (PADL) as MAP estimation with auxiliary latent variables, deriving closed-form bounds for the sparsity-storage-accuracy tradeoff and a data-driven activation threshold that eliminates grid search while boosting visual reconstruction and VLM inference compression.

Parameter symmetries determine representational geometry in overparameterized nonlinear networks

For one-hidden-layer nonlinear networks, this paper reduces parameter symmetries preserving the global function to feature addition, duplication, and scaling, proves that they can substantially reshape representational geometry, and gives sufficient conditions for minimum-norm selection to restore geometric identifiability.

PatchBench: Measuring Collateral Damage in Activation Patching

Addressing the blind spots of aggregate benchmarks that mask localized side effects, this paper introduces PatchBench—a curated bank of 400 verified, actionable jailbreak failures across 8 open-source LLMs—along with PatchBench-Local, an evaluation protocol that jointly measures harmful-neighbour correction and benign-neighbour preservation, revealing that state-of-the-art activation steering methods cause severe local collateral damage even when global MMLU remains virtually unchanged.

persona vectors: monitoring and controlling character traits in language models

This paper introduces an automated pipeline to extract linear "persona vectors" from language model activation space given only natural-language trait descriptions, demonstrating that these vectors can monitor and causally control character shifts during deployment, eliminate finetuning-induced persona drift without capability degradation via novel "preventative steering," and proactively screen problematic training data via projection differences prior to training.

PersonaManifold: Revealing and Exploiting Curved Geometry in LLM Persona Representations

PersonaManifold models LLM persona activations as a curved, anisotropic low-dimensional manifold, using graph shortest paths weighted by local metrics to measure similarity and interpolate personas; combined BST triplet consistency improves over Euclidean distance by 5.1–6.1 percentage points across three 7–8B models, primarily benefiting high-deviation persona pairs.

Phase Transitions in Attention: A Bayesian Theory of Copy Head Emergence

This paper develops a statistical field theory of Bayesian feature learning in attention networks that reduces the attention posterior to a low-dimensional order parameter space, demonstrating from first principles that softmax attention undergoes a first-order phase transition leading to copy head emergence, while linear attention exhibits a second-order phase transition followed by a smooth crossover.

Positional versus Symbolic Attention Heads: Learning Dynamics, RoPE Geometry, and Length Generalization

This paper investigates the learning dynamics of positional versus symbolic attention heads in RoPE-based Transformers on multi-hop reasoning tasks, proves constructive 2D geometric implementations, and introduces discrepancy bounds explaining why symbolic mechanisms extrapolate reliably while positional ones deteriorate rapidly.

PULSE: Identifying Demonstration-Utility Features with Sparse Autoencoders

PULSE identifies SAE features associated with demonstration utility from a small labeled discovery set, then uses them for complete-set ranking and cacheable per-example retrieval, improving selection across six text tasks while providing associational rather than causal-intervention evidence.

Reading Between the Dots: Decoding Hidden Computation across Filler Tokens

This paper investigates how frontier open-weights LLMs perform multi-step reasoning over semantically meaningless filler tokens (e.g., dots or counting sequences) where behavioral chain-of-thought monitoring fails, establishes that intermediate computational states are causally encoded in the residual stream, and introduces an unsupervised decoding pipeline that recovers intermediate values with 82%–94% accuracy without training or ground-truth supervision.

Rep2Text: Decoding Full Text from a Single LLM Token Representation

Rep2Text presents an end-to-end framework that decodes full input text from a single last-token representation of an LLM by mapping it through a lightweight gated residual adapter into a decoding LLM's embedding space, demonstrating that a single compressed activation retains roughly half of the original tokens without requiring iterative optimization.

Rethinking SAE Evaluation: An Atomic Interpretable Unit Framework Reveals Hidden Polysemanticity and Redundancy

This paper demonstrates that existing sparse autoencoder (SAE) metrics lack a unifying foundational objective, establishes an Atomic Interpretable Unit (AIU) basis paradigm with five derived necessary criteria, and introduces the Dominance-Aware Monosemanticity Evaluation (DAME) along with decoder redundancy metrics, uncovering critical failure modes where dominant features are systematically more polysemantic and redundant.

Root Cause Analysis of Measurement and Mechanistic Anomalies

This paper explicitly separates anomalies into latent mechanism shifts and observed measurement errors within structural causal models, proves generic identifiability for reduced intervention probabilities, and introduces Cali, an unsupervised algorithm achieving state-of-the-art root cause localization and anomaly-type classification.

Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders

This paper uncovers cross-modal feature heterogeneity where identical semantic concepts span different geometric directions across image and text embeddings, proving it drives the modality split in multimodal sparse autoencoders, and proposes decoupled modality-specific autoencoder training followed by Hungarian post-hoc latent alignment to achieve lossless reconstruction and superior cross-modal retrieval and concept steering.

Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?

Through synthetic experiments with known ground truth and three randomized frozen baselines (Frozen/Soft-Frozen Decoder and Frozen Encoder), this paper demonstrates that SAEs recover only 9% of true features despite 71% explained variance, and fully-trained SAEs barely outperform random frozen baselines in interpretability, sparse probing, and causal editing.

Semantic Concept Steering Breaks the Explanation Drift Loop in Continual Learning

Addressing the structural flaw where existing explanation-aware continual learning relies on self-referential frozen saliency targets that reinforce spurious shortcuts, this paper proposes Concept-Consistent Continual Learning (C3L), which anchors regularization to model-independent semantic concepts and leverages relevance conservation to crowd out background drift.

Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression

Replaces single decoder vectors with activation-weighted empirical probability distributions, leveraging optimal transport Wasserstein distance in a shared reference geometry to achieve state-of-the-art cross-layer SAE feature matching and automated feature circuit compression into interpretable supernodes.

Sink vs. Diagonal Patterns as Mechanisms for Attention Switch and Oversmoothing Prevention

This paper investigates the emergence of attention sinks versus diagonal attention patterns from geometric and parameter regularization perspectives, proving that sinks are the unique mechanism for hard attention switches with zero output and incur an asymptotically lower nuclear norm cost (\(O(T)\) vs. \(\Omega(T^{3/2})\)) than diagonal patterns under soft attention switches.

Sketching the Readout of Large Language Models for Scalable Data Attribution and Valuation

RISE leverages gradient energy concentration and exact outer-product factorization at the LLM readout layer (LM Head) alongside a dual-channel error formulation and factorized CountSketch compression, enabling forward-only, highly scalable data attribution and prospective valuation up to 32B parameters.

SMILE: Bridging Continuous Optimization and Discrete Symbolic Recovery

SMILE identifies decomposable structure in data, fits a network with fixed symbolic activations, and recovers compact expressions through pruning, constant refitting, and gradient-based rounding, achieving strong symbolic recovery under high noise in SRBench without universally leading noise-free recovery or prediction accuracy.

Sparse Reward Subsystem in Large Language Models

The paper reveals that hidden states in large language models host a sparse "reward subsystem" comprising less than 1% of neurons—including "value neurons" encoding state expected value and "dopamine neurons" encoding temporal difference (TD) errors—which causally govern reasoning and directly empower pre-generation confidence estimation and inference-time PRM search.

Spectral-Spatial Interpretation

To address the fundamental inability of 2D spatial heatmaps or global channel rankings to explain localized physical evidence in multispectral vision models, Spectral-Spatial Interpretation (SSI) formulates post-hoc explanation as local conditional games over position-channel players under fixed observed context, producing full-grid joint attribution matrices that strictly satisfy contextual aggregation consistency.

Task Vector Geometry Underlies Dual Modes of Task Inference in Transformers

By formalizing in-context learning via latent-variable sequence generation, this paper proves that transformers host two coexisting task inference modes: in-distribution tasks are retrieved via convex combinations of task vectors aligning with Bayesian posteriors, while out-of-distribution tasks are learned via a near-orthogonal subspace that linearly encodes running context statistics.

The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network

By extending the logit lens to the Post-LN architecture of Leela Chess Zero, this paper reveals that tactical puzzle solutions computed by internal look-ahead circuits are systematically overridden in final layers by learned safety priors; gradient-based activation steering against conservative preferences successfully recovers 61.7% of these forgotten puzzles.

The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching

Re-deriving activation patching from causal mediation analysis reveals that the Natural Indirect Effect (NIE) inherently bundles a mediator-bypass interaction term (INT) alongside the pure indirect effect, mathematically resolving long-standing anomalies including backup head omission, ranking divergence, and prompt-level faithfulness instability.

The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations

The paper reveals that temporal knowledge drift in LLMs is encoded along a geometric direction in the residual stream strictly orthogonal to both correctness and uncertainty, explaining why confidence-based and truthfulness-probing methods fail on outdated facts while enabling robust linear drift detection and causal steering.

The Rate-Distortion-Polysemanticity Tradeoff in SAEs

This paper introduces the Rate-Distortion-Polysemanticity (RDP) tradeoff in Sparse Autoencoders (SAEs), proving theoretically and empirically that enforcing monosemanticity incurs an unavoidable tax on rate or distortion, and that polysemanticity is an optimal response to concept co-occurrence in the data-generating process.

The Shape of Events: Edge-Based Inductive Biases via Cross-Domain Distillation

The paper uses standard knowledge distillation as an analytical tool to transfer an event teacher's predictions into an RGB student, finding joint improvements in tolerance to color changes, shape preference, and high-frequency-band noise, but with lower clean accuracy and sensitivity to disrupted geometric continuity rather than comprehensive robustness.

The SuperActivator Mechanism: Transformers Concentrate Reliable Concept Signals in the Tail

Reveals that concept-aligned attention heads multiplicatively amplify activation gaps and concentrate reliable semantics into an extreme right tail of "SuperActivators," where using only the top 5%–10% of activations improves multimodal concept detection F1 by up to 0.14 and substantially enhances attribution localization and faithfulness.

TPO: Tri-level Distributionally Robust Learning for OOD Direct Preference Optimization

Addressing multi-level distribution shifts in out-of-distribution (OOD) Direct Preference Optimization (DPO), TPO proposes a tri-level distributionally robust learning framework using hierarchically structured rectangular Wasserstein ambiguity sets, deriving an exact closed-form margin penalty, prompt embedding adversarial perturbation, and worst-case group reweighting with finite-sample convergence guarantees.

Training Language Models to Explain Their Own Computations

This paper introduces introspective interpretability by fine-tuning language models on supervision from automated interpretability tools, demonstrating that models leverage privileged access to explain their own internal features, causal activations, and decision rules significantly better and with roughly 100x greater sample efficiency than external or larger models.

Transformers Linearly Represent Highly Structured World Models

Through mechanistic interpretability of an 8-layer transformer trained on Sudoku solving traces, this paper demonstrates that internal representations do not follow physical grid cells but organize around 27 constraint substructures (rows, columns, boxes), culminating in a sparse circuit of 91 monosemantic neurons in the final MLP that commit to naked-single decisions.

Understanding the Interplay between Memorization and Learning in Large Language Models

This paper introduces Contextual Memorization to disentangle rote memorization from generalizable contextual learning, proves through controlled formal-language testbeds that optimal language learning inevitably entails local memorization, and develops Memorization-Aware Training (MAT) to dynamically down-sample learned strings for more uniform and stable generalization.

What Comes Next and Why: Interpretable Next-Event Prediction with Neuro-Symbolic Rules

Addressing the trade-off between black-box sequence models and uncomposable symbolic pattern miners in next-event prediction, this paper introduces Rules and Runes, which differentiably learn temporal conjunctive rules via two-phase temperature annealing and combine them with an additive residual network to rival Transformer accuracy while preserving transparent rule-level explanations.

Witness Overlap: Directional Provenance Inside Open-Weight Model Families

Within known same-family open-weight models with aligned parameters, Witness Overlap introduces a third checkpoint, compares each candidate endpoint's update overlap toward the target and witness, and predicts the lower-overlap endpoint as the parent; Frobenius cosine achieves 95.3% orientation accuracy over 1,542 single-witness triplets from 16 LLM families.

Explainability matters: The effect of liability rules on the healthcare sector

From an interdisciplinary perspective bridging tort law and game theory, this paper analyzes two polar liability models—the black-box "Oracle" and the transparent "AI Colleague"—demonstrating that high explainability is not merely a clinical convenience but a pivotal legal safeguard preventing defensive medicine driven by uncritical compliance with AI.