Search: “mechanistic-interpretability”

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?

SAEScientist-Bench evaluates whether AI agents can autonomously conduct SAE interpretability research in Gemma-2-9B-IT, finding frontier agents trail expert baselines.

SAEScientist-Bench tests if AI agents can act as scientists using SAE tools for autonomous mechanistic discovery, requiring them to design contrastive probes and navigate a Gemma Scope dictionary of 131K+ features in Gemma-2-9B-IT. Across 10 agent configurations and 20 tasks, frontier agents showed genuine discovery capability but remained well behind expert reference features, lagging most in causal steering. Agents frequently misinterpreted experimental measurements even when designing effective contrasts.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI research

55

Explainability Assistant: A Conversational XAI Interface for Interpreting Energy Consumption Models

Researchers release Explainability Assistant, an open-source conversational XAI tool using LLM function calling, lifting intent-parsing accuracy from 76.8% to 94%.

The paper introduces the Explainability Assistant, an open-source conversational XAI system for interpreting energy consumption forecasting models such as genetic-programming symbolic regressors. It uses LLM function calling instead of rigid custom grammars, achieving 94% intent-parsing accuracy versus 76.8% for prior work TalkToModel, and adapts to different ML problem types without task-specific fine-tuning. Comparative evaluation with energy domain specialists against a traditional XAI dashboard showed improved usability, with all experts preferring the conversational interface.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:40 UTC · 5d agoAI research1

25

SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?

SAEScientist-Bench tests whether AI agents can autonomously run SAE interpretability research on Gemma-2-9B-IT; frontier agents trail expert baselines.

The benchmark requires agents to design contrastive probes and navigate the Gemma Scope dictionary of over 131K features in Gemma-2-9B-IT to discover optimal interpretable features, scored against expert-curated references on Neuronpedia via activation rank, concept selectivity, and causal steering. Across 10 agent configurations and 20 tasks, frontier agents demonstrate genuine discovery capability and approach expert levels at separating target concepts from controls, but lag substantially in causal steering and frequently misinterpret experimental measurements. The authors frame this as establishing experimental model understanding as a measurable capability for closed-loop autonomous AI R&D and post-hoc monitoring for recursive self-improvement.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:45 UTC · 7d agoAI research1

40

From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge

Interpretability study traces how Qwen, Llama, and Gemma route query information and internal knowledge across layers when answering questions.

Researchers used layerwise interventions on hidden states to separate query-routing signals from target knowledge in language models, testing Qwen, Llama, and Gemma on country-continent questions with varied answer types. A pair-conditioned request direction strengthens before interventions alter downstream knowledge, opening a causal window while answer-supporting content is still forming. Trajectories differ by model: Gemma shows a partially overlapping mid-layer routing profile, while Llama has no sustained routing-effect window under the same gates.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:39 UTC · 5d agoAI research1

25

Can your coding style predict whether your code is vulnerable?

University of Massachusetts Dartmouth researchers present VulStyle, a stylometry-based vulnerability detector that also exposes benchmark reliability problems.

VulStyle combines stylometric features with syntax-tree structure and source tokens, pre-trained on about 4.9 million functions across seven programming languages and fine-tuned on five vulnerability detection datasets. It beat token-only detectors on some benchmarks but its F1 drops sharply on DiverseVul, which the authors link to noisy labels inflating reported performance across popular datasets. The authors argue style-aware detection should be harder to evade but did not test this empirically, and they note that uniform LLM-generated code may strip away the individual developer style the model depends on.

Help Net Security·Aug 24, 04:21 UTC · 23d agoResearch1

48

Augustinian BabyLM: What Ostensive Definition Can and Cannot Teach a Small Language Model

Study shows visually grounded token embeddings in a small masked LM persist through training and improve object-property knowledge, but escape standard BabyLM benchmarks.

The paper implements ostensive definition for a small DeBERTa masked language model trained on 10M words, seeding visually grounded tokens with embeddings derived from labeled image regions before training. Visual initialization leaves a persistent, seed-replicated advantage on object-property knowledge (COMPS) and a corpus-tailored Visual-Property Swap benchmark covering color, material, size, and shape, but has no effect on most BabyLM grammar benchmarks. Synthetic grounding of previously unseeded words causally transfers the advantage to exactly those words.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:43 UTC · 5d agoAI research

20

Curriculum Learning as Transport: Understanding Curricula with Wasserstein Geodesics

Researchers model curriculum learning as Wasserstein transport over difficulty distributions, finding curriculum benefits are strongly task- and budget-dependent with no dominant strategy.

The framework represents curricula as trajectories of training distributions over discrete difficulty levels, decoupling ordering, matched exposure, endpoint smoothness, and pacing. Across a calibrated suite of 12 tasks and 33 difficulty axes under fixed training budgets, no single strategy dominates, though easy-to-hard ordering improves hard-level performance relative to exposure-matched static sampling. Endpoint smoothness and pacing substantially affect where along the difficulty spectrum a curriculum is effective, and the transport view supports extensions to learned pacing and structured difficulty spaces.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:39 UTC · 7d agoAI research1

28

Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions

A linear hidden-state direction encodes question impossibility in 1.7B-70B LLMs, but misalignment with the safety-refusal pathway explains why models answer unanswerable questions.

The study examines why instruction-tuned LLMs from 1.7B to 70B parameters answer structurally unanswerable math and code questions instead of abstaining. A single linear direction in the hidden state separates answerable from impossible prompts, showing models represent impossibility before generation, but this direction is nearly orthogonal to the canonical safety-refusal direction. Generation-time steering along the recognition direction changes invalidity-aware behavior dose-responsively, and the geometry is present even at the pretraining endpoint, indicating a routing failure rather than an encoding failure.

Hugging Face daily papers·Aug 28, 20:00 UTC · 18d agoAI safety & security

30

Thought without systematicity? Evaluating reasoning models on rule induction tasks

Study finds reasoning models often fail on structurally equivalent variants of tasks they solve, suggesting their reasoning lacks systematicity.

The paper extends rule induction tasks from cognitive science using task isomorphisms such as recombination and substitution to test systematicity in reasoning models. Despite solving tasks correctly, models frequently fail on structurally equivalent variants of the same task. The authors conclude many model behaviors lack systematicity, making it difficult to establish cognitive abilities beyond the specific evaluation contexts.

Hugging Face daily papers·Sep 11, 20:00 UTC · 4d agoAI research

35

A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models

Study shows video models often learn correct physics but fail to use it; low-dimensional 'causal writability' edits can restore correct motion.

The paper demonstrates 'causal writability' in video generation models: physically correct motion remains available inside the model even when the model outputs incorrect motion. In a red/blue mass oscillation setup, a low-dimensional edit predicted from simple physical variables restores correct fast motion, with a sharp depth boundary marking commitment. Early causal writability predicts which training errors later get corrected, and both writability and closure reproduce in a pretrained 1.3B video model.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:58 UTC · 1d agoAI research

33

The Semantic Bottleneck: Leveraging Semantic Representations for Non-Invasive Speech Decoding

Researchers introduced Brain2Semantics2Text, decoding sentence meaning from non-invasive MEG brain recordings via a semantic bottleneck, improving on prior Brain2Text methods.

The paper proposes Brain2Semantics2Text, a non-invasive speech decoding method that maps sentence-level magnetoencephalography (MEG) responses into a semantic embedding space and inverts those embeddings into natural language. Motivated by evidence that high-level semantic representations are distributed across cortex and evolve on slower timescales, the approach targets meaning rather than phonemes or words, avoiding the need for word-level alignment. The authors report improved sentence-level results compared to prior non-invasive Brain2Text methods despite the low signal-to-noise ratio of neural recordings.

Hugging Face daily papers·Sep 8, 20:00 UTC · 7d agoAI research2

25

Necessary or Sufficient? Evaluating LLM Explanations With Behavioural Evidence

New framework tests whether LLM-cited explanation factors are necessary or sufficient, finding weak correlation across Claude, GPT, and Gemini models.

An arXiv paper introduces black-box intervention tests measuring whether factors LLMs cite in their explanations are necessary or sufficient for their outputs in agent oversight workflows. Across eight models from the Claude, GPT, and Gemini families, Spearman correlations between cited rankings and measured influence ranged from 0.349-0.354 (advisor recommendation) to 0.431-0.580 (prompt monitoring). Uncited factors scored above the lowest cited factor in up to 57.6% of advisor responses, showing cited top-three factors do not reliably identify the most influential inputs.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:37 UTC · 11d agoAI safety & security1

44

Why Is SHAP Not a Reliable Standalone Explanation Framework for Malware Detection?

arXiv paper shows SHAP gives unreliable standalone explanations for malware detection, with attribution dilution and sign reversal in dependent PE feature spaces.

The paper argues SHAP's formal guarantees are insufficient for reliable malware interpretation because the explained feature-coalition game is fixed only by analyst choices, not by malware behavior in the data. In static Portable Executable feature spaces, dependent feature groups cause conditional SHAP to dilute credit by a factor of 1/m across redundant features, attribute importance to features the model never uses, and even reverse attribution signs; interventional SHAP queries off-manifold coalitions no real executable exhibits. Experiments on EMBER-2018, EMBER-2024, and BODMAS with fixed LightGBM and XGBoost detectors confirm these effects. The authors position SHAP as a limited diagnostic requiring explicit data-distribution statements and domain validation, not a standalone explanation framework.

arXiv cs.CR·Sep 4, 02:00 UTC · 12d agoResearch1

44

LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence

LimiX-2 scales Contextual Mechanism Networks pretrained via context-conditional masked modeling, beating tabular foundation models on TabArena, TALENT, and BCCO benchmarks.

LimiX-2 is a new tabular model in the LimiX family, developed through model and data scaling guided by previously established scaling laws under the Contextual Mechanism Networks (CMNs) paradigm. It is pretrained with Context-Conditional Masked Modeling (CCMM) on synthetic datasets generated by structural causal models spanning diverse graph structures, functional mechanisms, and observation processes. It outperforms dataset-specific models and tabular foundation models on TabArena, TALENT, and BCCO, and its feature attention encodes direct causal relationships, enabling accurate causal skeleton recovery.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:30 UTC · 18h agoAI research1

18

Type Diversity Enables Transformers to Generalise Compositionally

Researchers show lexical-versus-structural compositional generalization gaps in Transformers stem from type diversity imbalance in datasets, not architectural limits.

The paper argues that Transformers' difficulty with structural compositional generalization is an artifact of low structural type diversity in prior benchmark datasets rather than an architectural limitation. Using Grammatical Framework, the authors create linguistically diverse variants of COGS and SLOG. They find type diversity correlates with compositional generalization equally in lexical and structural test cases, contradicting previous claims that compound divergence explains task difficulty.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 17:59 UTC · 4d agoAI research

20

Technical Manual for a Toolkit for Measuring Contextual Individuation in Transformer Language Models

An open methodology toolkit measures whether transformer language models contextualize fixed word forms across domains using bridge forms and layer-wise silhouette analysis.

The manual documents an open toolkit built around 'bridge forms' - identical written words recurring across two or more subject domains with a different sense in each - to test whether transformer language models individuate word occurrences by context beyond the embedding layer. It covers declarative specification of bridge forms, Wikipedia corpus acquisition, occurrence localization, layer-wise representation extraction, domain-pairwise silhouette measurement, and visualization, justifying each choice against failure modes such as sense contamination and subword-tokenization misalignment. It is a methodological and implementation reference and reports no empirical results.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 16:38 UTC · 11d agoAI research

14

Observational Indistinguishability and Integrity Blind Regions in Hybrid Quantum-Classical Workflows

Framework formalizes integrity blind regions in hybrid quantum-classical workflows, validated across 3,600 label interventions with conformal detection rules.

The paper presents a claim-relative evidence and reference framework for integrity of hybrid quantum-classical workflows, distinguishing structural blind regions caused by observational indistinguishability from finite-batch statistical misses. Experiments over 3,600 label interventions show exact label-path invariance for feature and prediction views. The geometry-aligned construction detects 343 of 2,700 conclusion-changing interventions using the conformal rule and 1,183 of 2,700 with the uncorrected union, with executed conformal clean false-action rates of 0.048-0.059.

arXiv cs.CR·Sep 15, 13:16 UTC · 22h agoResearch

12

Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs

Study shows LLM reasoning operations like planning and deduction are geometrically separable in hidden states, with separability peaking in middle layers.

Researchers investigate whether functional reasoning operations — problem formulation, goal decomposition, deduction — have corresponding geometric structure in LLM hidden representations. They find operations are separable in held-out representations with separability peaking in middle layers, ruling out lexical and positional confounds; token-wise operation alignment becomes more distributed across layers, and identical surface tokens are represented differently depending on their surrounding chunk. Attention-masking interventions show chunk-onset operation-aligned representations depend on preceding reasoning context; code is released on GitHub (naver-ai/beneath-cot).

Hugging Face daily papers·Sep 3, 20:00 UTC · 12d agoAI research1

40

General Quantification of Covariate and Concept Shifts

Paper proposes γ*-concept shifts via entropic optimal transport, deriving estimable generalization bounds unifying covariate and concept shift under distribution shift.

The authors show existing definitions of concept shift break when source and target supports mismatch and propose γ*-concept shifts grounded in entropic optimal transport. They derive a general error bound covering broad loss functions, label spaces and stochastic labeling, plus estimators with concentration guarantees. The resulting DataShifts algorithm quantifies distribution shifts and estimates the error bound in most applications, addressing learning bounds that were previously non-estimable from samples.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:57 UTC · 5d agoAI research1

28

One Symptom, Three Levers: A Critical Review of On-Policy Self-Distillation

A review paper frames on-policy self-distillation collapse as governed by three levers: token weighting, privileged information, and guidance decay.

The paper critically reviews On-Policy Self-Distillation (OPSD), where a language model trains on its own generations scored token-by-token by a teacher conditioned on privileged information such as reference solutions or environment feedback. It identifies collapse, the progressive narrowing of producible reasoning paths, as the dominant failure mode and analyzes it through three levers: signal weighting, the nature of privileged information, and teacher dynamics. The review is restricted to mathematical reasoning, reports no new experiments, and offers a shared vocabulary separating settled findings from disputed ones.

Hugging Face daily papers·Aug 25, 20:00 UTC · 21d agoAI research

45

Molecular Déjà Vu: Digit-Level Retrieval of Published Values in Frontier Language Models

Audit of 22 frontier models finds widespread verbatim retrieval of published molecular property values, with higher reasoning increasing recall of memorized numbers.

An arXiv audit tests 22 frontier LLMs across 12 molecular regression benchmarks for verbatim retrieval of published values. More than 50% of the LLMs show verbatim retrieval on five datasets, and identical experiments are flagged 89% more often at a high reasoning level than at the lowest one. Suppressing retrieval moves model prediction errors closer together in relative terms, suggesting predictive capability is not determined solely by memorized values.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:32 UTC · 11d agoAI research1

42

Large Language Models Develop Belief State Geometry In-Context

Probing six open-source LLMs on HMM-generated data shows belief states are linearly decodable (R² 0.83–0.99), suggesting in-context learning approximates Bayesian prediction.

Researchers prompted six open-source LLMs with data from 40 hidden Markov models selected for non-trivial belief structure and probed residual-stream activations for belief states (posteriors over hidden states). Belief states were linearly decodable with peak R² values of 0.83–0.99 across HMM/LLM combinations, spanning early to late layers. Patching and steering the probe-identified subspace preserved downstream prediction quality while control interventions degraded performance substantially, establishing functional relevance. The results provide representation-level evidence that in-context learning approximates optimal Bayesian prediction over a context-inferred generative model.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 16:17 UTC · 19h agoAI research

45

A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM

A*-Thought-V2 compresses chain-of-thought into latent tokens using geometric hidden-state dynamics, cutting computation while improving accuracy on Qwen models.

A*-Thought-V2 models chain-of-thought as a hidden-state trajectory and interleaves explicit text with continuous latent tokens, compressing steps whose transitions deviate from the question-to-solution direction. Trained via stepwise embedding forcing and label forcing with soft multi-modal supervision, it was evaluated on Qwen3.5-9B and Qwen3.6-27B across six benchmarks. Reported results include up to 2.6% average accuracy gain, up to 50% shorter responses, 2.29x higher Accuracy per Computation Unit, 94.6% faster preprocessing, and up to 80.3% faster training.

Hugging Face daily papers·Sep 7, 13:56 UTC · 8d agoAI research

48

Perturbation Probing: A New Diagnostic for the Fragility of LLM Safety

Unit 42 research shows LLM safety refusals concentrate in a thin neural layer, motivating external, multi-layered AI security controls.

Palo Alto Networks Unit 42 introduces Perturbation Probing, a diagnostic technique for measuring the fragility of LLM safety mechanisms. The research finds that safety refusal behavior is localized within a thin neural layer, implying small perturbations can undermine built-in refusals. The authors argue this motivates external, multi-layered security defenses on top of model-internal safety training.

Palo Alto Unit 42·Aug 28, 22:00 UTC · 18d agoAI safety & security

45

MemSentry: A Framework for Detecting Persistent Memory Poisoning in Agentic AI

MemSentry intercepts persistent-memory writes in agentic AI to catch memory poisoning, reaching 91.7% accuracy with SBERT+LR classification.

Memory poisoning lets adversaries plant crafted content in an agent's long-term memory to suppress security alerts, enable privilege escalation, or override policies without modifying model weights or system prompts. The paper presents MemSentry, a configuration-driven framework that evaluates proposed persistent-memory writes on source trust, semantic risk, attack radius over a dependency DAG, access risk, and a signed security-state delta to issue deterministic Accept, Review, or Quarantine decisions. Across 1,000 GPT-4-generated scenarios on a 20-asset dependency DAG, SBERT+LR achieved 91.7% accuracy and 0.908 macro-F1, all four classifiers detected 100% of external quarantine-class threats, and verified-insider writes are escalated for human review rather than auto-quarantined.

arXiv cs.CR·Sep 8, 13:40 UTC · 7d agoAI safety & security

48

Bag of Tricks or Bag of Myths? Reducing Modeling Complexity with Task Knowledge in Explainable Suicide Risk Assessment

Researchers audit 31 NLP techniques for clinician-annotated suicide risk prediction, finding only 5 of 31 comparisons yield reliable gains.

A study of 1,635 clinician-annotated social media posts ran roughly 300 controlled experiments across 7 methodological families, auditing techniques such as model scaling, synthetic data, ensembling, and threshold tuning under severe class imbalance. The proposed system reformulates risk factor prediction as entailment between posts and codebook definitions, using architecturally diverse ensembles with class-balanced training and deployment-consistent calibration. It scores 0.8203 for risk, 0.7953 for evidence, and 0.7045 macro-F1 for factors, ranking third among 53 teams.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:14 UTC · 8d agoAI research

14

Understanding the Impact of Model Pruning on Long-Tail Forgetting and Explanation Reliability in Medical Imaging

Systematic study finds model pruning causes frequency-dependent long-tail forgetting in medical imaging and that gradient-informed methods best preserve explanations.

Across two long-tailed medical imaging datasets, two CNN architectures, four pruning methods, and sparsity up to 95%, the study measures predictive performance, explanation stability, and faithfulness. Rare classes degrade earlier and more severely than frequent ones, while explanation reliability depends mainly on the pruning strategy, with gradient-informed methods degrading least. Mechanistic analysis ties explanation collapse to loss of class-discriminative gradients rather than vanishing feature activations, recommending class- and explanation-aware evaluation of compression.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:44 UTC · 8d agoAI research

27

Mythos Vulnerability Firehose Hits a Human Bottleneck

Analysis of Project Glasswing findings shows only a fraction of discovered vulnerabilities have reached disclosure and even fewer are fixed.

Dark Reading reports that an analysis of Project Glasswing findings shows only a fraction of the vulnerabilities discovered by the program have reached disclosure, and an even smaller number have been fixed. The article examines how the volume of findings from the discovery program is bottlenecked by limited human triage and remediation capacity. The piece highlights growing tension between high-volume vulnerability discovery and the industry's ability to process, disclose and patch reports.

Dark Reading·Sep 9, 21:19 UTC · 6d agoResearch

40

Diffusion Models and Concept Formation

Paper argues diffusion models implicitly form Cobweb-like concept hierarchies, with a basic level emerging at intermediate noise levels.

The authors draw a formal correspondence between diffusion models and Cobweb, a classic incremental concept-hierarchy learner, noting both are hierarchical Bayesian density models with Gaussian prototypes. Modes of the diffusion model's noisy marginals form a hierarchy whose basic level sits at intermediate noise, where class identity commits. The correspondence is tested on MNIST and Fashion-MNIST via mode-finding. Diffusion is reframed as a cognitive model of concept formation.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 16:43 UTC · 4d agoAI research

15

DynSHAP: Towards Explainable Dynamic Survival Analysis

DynSHAP extends SHAP explainability to dynamic survival analysis, treating time-feature pairs as Shapley players for longitudinal clinical predictions.

DynSHAP adapts marginal SHAP estimators to dynamic survival analysis by treating time-feature pairs as players in the Shapley game, handling longitudinal irregular inputs and functional survival outputs. Temporal DynSHAP learns linear feature dependencies over time and addresses them with conditional sampling. On synthetic data with ground-truth attributions it recovers temporally dependent features more accurately than marginal estimators, and it produces faithful attributions on two real-world clinical datasets across two DSA architectures.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 16:39 UTC · 4d agoAI research

22