Searcharxiv⌕ Search

arXiv subjects

Shay B. Cohen

Publications and source records attributed to Shay B. Cohen.

At least 19 recordsLinked to original sources

Understanding AI from the Inside: The State of Mechanistic Interpretability for AI Safety

Neural networks are likely to be among the most consequential artefacts in human history. As these systems become more capable and autonomous, their potential to cause societal harm also grows, making AI safety one of the defining challenges of our time. Yet we continue to build increasingly powerful systems without fully understanding how they work, how they will behave in unfamiliar situations, or how to reliably control them. Mechanistic interpretability offers a path towards understanding these systems by studying their internal representations and computations, with the aim of evaluating, predicting, and controlling behaviour. We survey mechanistic interpretability for AI safety in order to answer the question: Does understanding how models work actually make them safer? We place mechanistic interpretability in the broader context of interpretability research and compare its mechanistic evidence with behavioural evaluation. We introduce the foundations of mechanistic interpretability, including features, polysemanticity, transformer computation, and circuits, and survey methods for feature extraction and analysis, circuit discovery and causal verification, explaining intermediate representations in natural language, learning dynamics, and training-data attribution. Beyond surveying existing methods, we propose a unified framework for evaluating mechanistic interpretability claims and characterise six research orientations by their goals, assumptions, and trade-offs. Finally, we examine how mechanistic understanding can support AI safety, including knowledge editing, steering, and monitoring, and assess its limitations, failure modes, open problems, and relationship to behavioural safety measures. We also release a repository of code notebooks with worked examples that reproduce key findings from surveyed methods, allowing readers to build intuition through hands-on experimentation.

cs.CL↗

They Can See It but Not Say It: Iterative Self Knowledge Re-expression in Visual Reasoning Relative Pose Identification

The task of visual reasoning relative pose identification (VRRPI) has exposed critical weaknesses in the multi-view 3D spatial reasoning capabilities of modern vision-language models (VLMs). We present a systematic analysis of VLMs' behavior in VRRPI, uncovering stable output biases where VLMs disproportionately predict specific motion patterns regardless of visual evidence. To mitigate these inherent biases, we propose self knowledge re-expression with iterative debiasing (SKR-ID), an annotation-free pipeline that adapts VLMs for VRRPI using only randomly extracted unannotated image pairs. SKR-ID transitions VLMs' output mechanism from generic next-token prediction to task-optimized classification heads. The pipeline proceeds in two stages: (1) a cold-start initialization using debiasing object-centric prompts to anchor motion inference in 2D visual evidence; and (2) a recursive logit-ranking mechanism that enforces statistical parity by using the median logit value as a dynamic threshold. Experimental evaluation across frontier VLMs on standard benchmarks demonstrates that SKR-ID significantly mitigates systematic biases. Our method improves open-source VLMs' Macro-F1 scores from 35.0-45.4% to 53.7-58.0%, without using any annotated data. These results show that SKR-ID can effectively improve 3D reasoning performance from a severely collapsed baseline. The code is publicly available at https://github.com/astudentuser/Codes_of_SKR-ID.

cs.CV↗

Transferrable Surrogates in Expressive Neural Architecture Search Spaces

Neural architecture search (NAS) faces a challenge in balancing the exploration of expressive, broad search spaces that enable architectural innovation with the need for efficient evaluation of architectures to effectively search such spaces. We investigate surrogate model training for improving search in highly expressive NAS search spaces based on context-free grammars. We show that i) surrogate models trained either using zero-cost-proxy metrics and neural graph features (GRAF) or by fine-tuning an off-the-shelf LM have high predictive power for the performance of architectures both within and across datasets, ii) these surrogates can be used to filter out bad architectures when searching on novel datasets, thereby significantly speeding up search and achieving better final performances, and iii) the surrogates can be further used directly as the search objective for huge speed-ups.

cs.LG↗

Seeing the Forest for the Trees: Emergent Agreement from Individual Conversations

As LLM agents interact in the wild and shape each other's decisions, understanding how they converge toward agreement becomes important. We study this in multi-turn conversations where two LLMs discuss subjective matters, can agree or disagree, and may change their stances. Using a linear classifier on individual conversation activations, we identify an internal representation of agreement. We find that it varies substantially across turns and conversations and does not reliably track external agreement at the individual conversation level. We therefore move beyond the representation of agreement in individual conversations, by aggregating the agreement representation across conversations. Externally, this aggregate view reveals a simple structure: when pooled across conversations, agreement can be predicted well with a Markov process, with agreement at each turn being a direct function of agreement at the previous turn. Similarly, when internal agreement is aggregated, it couples with the external Markov agreement through a power-law mapping, exposing Markovian dynamics in internal representations. Additionally, we find that after a change of stance, models express agreement externally before their internal representations reflect the same shift. These results reveal a joint structure between internal and external agreement that emerges only when internal features are aggregated across conversations. Code available at https://github.com/technion-cs-nlp/SeeingTheForestForTheTrees.

cs.CL↗

Beyond Context Length: Characterising the Difficulty of Long-Context Reasoning

Long-context language models support ever longer context windows, and evaluations typically treat context length as the main indicator of difficulty. However, context length alone cannot explain why models fail on long-context tasks that require reasoning over complex information. Contexts of equal length can pose very different challenges: a task may be solvable from a single local passage, or it may require integrating evidence distributed across the full context. We quantitatively characterise this difficulty through three properties: evidence scope, how much necessary information must be found; evidence dispersion, how widely that information is scattered within the context; and reasoning depth, how many dependent steps are needed to derive an answer from it. We construct LURCH, a multi-domain question-answering benchmark of 6,550 instances with supporting evidence and reasoning graphs, enabling all three properties to be quantified. Evaluations of proprietary and open-weight models show that: (1) at fixed context length, all our three properties are strongly negatively correlated with model performance; (2) as context length increases, performance degrades sharply on samples with dispersed evidence, whereas it remains relatively stable on samples with localised evidence; and (3) difficulty compounds across properties, as questions with moderately high values on all three are harder than those with an extreme value on just one. These findings establish evidence scope, dispersion and reasoning depth as measures of long-context reasoning difficulty that complement context length.

cs.CL↗

Are you Synthesizing or Recalling? Evaluating LLMs on Algorithmic Code Retrieval

Large language models (LLMs) have demonstrated strong performance in code generation, where success depends on both recalling relevant algorithmic knowledge and reasoning about how to apply it. However, existing LLM pipelines are opaque, with no explicit separation between these two components. We argue that for well-known algorithms whose canonical implementations are widely accessible in pretraining corpora, code generation is better measured as \textit{parametric code retrieval}: reproducing a named algorithm from internalised knowledge rather than synthesizing a novel one. We introduce AlgoREval, a benchmark of 599 problems spanning classical 77 algorithms across 14 domains, 7 programming languages, and 4 graph-input representations to evaluate this capability in isolation, and assess 15 models (7B--34B parameters) in a zero-shot setting. We find substantial variation in retrieval accuracy across languages and input representations, even for widely documented algorithms and show that prompt augmentation with retrieved code snippets or structured algorithmic hints improve accuracy on complex algorithms, while SFT achieves broader language gains and GRPO achieves larger per-language gains on specific languages. Together, our results establish parametric code retrieval as a distinct, measurable capability and caution against deploying AI-generated algorithmic code without systematic validation.\footnote{Code and dataset are available at https://github.com/Nickil21/AlgoREval

cs.LG↗

When Can Attention Heads Be Statically Defined?

Some attention heads learn similar patterns across inputs. Reusing these patterns could reduce training cost by avoiding repeated query-key score computation and softmax. Through controlled pretraining comparisons, we identify Selective Attention Freezing (SAF), which selects heads with low attention-pattern variance and replaces their attention weights with fitted post-softmax means halfway through training. We represent these fixed patterns with absolute-position and relative-distance preferences, reducing storage from quadratic to linear in sequence length. A fused kernel reconstructs the patterns and executes ordinary-attention and replaced heads together. At matched training-token budgets, replacing 25% of attention heads gives 1.056x faster post-replacement optimiser updates at 124M parameters and 4K context, with a 0.77% perplexity increase. At 1B and 8K context, post-replacement updates are 1.068x faster on four GPUs including communication, with a 0.51% perplexity increase. The resulting models also accelerate long-input finetuning and causal prefill. After associative-recall adaptation, the 124M model with 25% replacement generalises to more key-value pairs at a fixed length better than ordinary attention and two pruning controls.

cs.LG↗

ONNX-Net: Towards Universal Representations and Instant Performance Prediction for Neural Architectures

Neural architecture search (NAS) automates the design process of high-performing architectures, but remains bottlenecked by expensive performance evaluation. Most existing studies that achieve faster evaluation are mostly tied to cell-based search spaces and graph encodings tailored to those individual search spaces, limiting their flexibility and scalability when applied to more expressive search spaces. In this work, we aim to close the gap of individual search space restrictions and search space dependent network representations. We present ONNX-Bench, a benchmark consisting of a collection of neural networks in a unified format based on ONNX files. ONNX-Bench includes all open-source NAS-bench-based neural networks, resulting in a total size of more than 600k {architecture, accuracy} pairs. This benchmark allows creating a shared neural network representation, ONNX-Net, able to represent any neural architecture using natural language descriptions acting as an input to a performance predictor. This text-based encoding can accommodate arbitrary layer types, operation parameters, and heterogeneous topologies, enabling a single surrogate to generalise across all neural architectures rather than being confined to cell-based search spaces. Experiments show strong zero-shot performance across disparate search spaces using only a small amount of pretraining samples, enabling the unprecedented ability to evaluate any neural network architecture instantly.

cs.LG↗

MoRFI: Monotonic Sparse Autoencoder Feature Identification

Large language models (LLMs) acquire most of their factual knowledge during the pre-training stage, through next token prediction. Subsequent stages of post-training often introduce new facts outwith the parametric knowledge, giving rise to hallucinations. While it has been demonstrated that supervised fine-tuning (SFT) on new knowledge may exacerbate the problem, the underlying mechanisms are still poorly understood. We conduct a controlled fine-tuning experiment, focusing on closed-book QA, and identify latent directions causally implicated in this degradation. Specifically, we fine-tune Llama 3.1 8B, Gemma 2 9B and Mistral 7B v03 on seven controlled mixtures of a single QA dataset, controlling for the percentage of new knowledge and number of training epochs. By measuring performance on the test set, we validate that incrementally introducing new knowledge increases hallucinations, with the effect being more pronounced with prolonged training. We leverage pre-trained sparse autoencoders (SAEs) to analyze residual stream activations across various checkpoints for each model and propose Monotonic Relationship Feature Identification (MoRFI) for capturing causally relevant latents. MoRFI filters SAE features that respond monotonically to controlled fine-tuning data mixtures of a target property. Our findings are consistent with exposure to unknown facts disrupting the model's ability to retrieve stored knowledge along a set of directions in the residual stream. Our pipeline reliably discovers them across distinct models, partially recovering lost knowledge through single-latent interventions.

cs.CL↗

PiCSAR: Probabilistic Confidence Selection And Ranking for Reasoning Chains

Best-of-n sampling improves the accuracy of large language models (LLMs) and large reasoning models (LRMs) by generating multiple candidate solutions and selecting the one with the highest reward. The key challenge for reasoning tasks is designing a scoring function that can identify correct reasoning chains without access to ground-truth answers. We propose Probabilistic Confidence Selection And Ranking (PiCSAR): a simple, training-free method that scores each candidate generation using the joint log-likelihood of the reasoning and final answer. The joint log-likelihood of the reasoning and final answer naturally decomposes into reasoning confidence and answer confidence. PiCSAR achieves substantial gains across diverse benchmarks (+10.18 on MATH500, +9.81 on AIME2025), outperforming baselines with at least 2x fewer samples in 16 out of 20 comparisons. Our analysis reveals that correct reasoning chains exhibit significantly higher reasoning and answer confidence, justifying the effectiveness of PiCSAR.

cs.CL↗

Debiasing Without Protected Attributes: Latent Concept Erasure from Textual Profiles

Most fairness research in NLP assumes direct access to protected attributes such as gender, race, or nationality. In practice, however, such information is often unavailable due to privacy constraints, missing metadata, or legal restrictions, even though models may infer it from indirect textual cues. This raises a key question: can debiasing succeed without direct access to sensitive attributes? We propose H-SAL, which performs post-hoc concept and attribute erasure using self-description text as an implicit debiasing signal. To support this setting, we introduce a multi-domain Stack Exchange-based fairness benchmark for helpfulness prediction that includes both explicit and implicit signals, enabling comparison between standard debiasing with protected labels and debiasing without access to sensitive information. Across encoder and decoder-only language models, we find that implicit self-description often matches or outperforms explicit-label-based debiasing. Our results broaden representation-level fairness research and provide a new benchmark for studying debiasing under realistic data constraints.

cs.CL↗

Pythagoras-Prover: Advancing Efficient Formal Proving via Augmented Lean Formalisation

Modern Lean theorem provers achieve strong performance only with substantial training and inference compute, driven in part by scarce verified proof data and the long reasoning traces of formal proof search, making both supervised fine-tuning (SFT) and sampling expensive. We introduce Pythagoras-Prover, a compute-efficient open-source family of Lean theorem provers built for practical compute budgets. The family spans two generation paradigms: autoregressive models at 4B and 32B parameters, and a first proof-of-concept diffusion-based prover (4B) that iteratively refines Lean proofs at inference time. For training efficiency, we build a Lean-verified corpus stratified into easy, medium, and hard problems for curriculum SFT, so models acquire proof skills progressively from shorter, simpler proofs to longer, harder ones. During SFT, a dynamic proof-reasoning filtering scheme preserves informative proof traces while keeping each instance within an 8k-token context budget. We also introduce Augmented Lean Formalisation (ALF), which expands scarce verified corpora into variants of formal statements, populated via self-distillation for extra training signal without formally verifying every mutated instance. By perturbing known problems while preserving their formal character, ALF reduces reliance on any statement's surface form. Empirically, Pythagoras-Prover-4B surpasses DeepSeek-Prover-V2-671B at pass@32 on MiniF2F-Test (86.1% vs 82.4%) with ~167x fewer parameters, while Pythagoras-Prover-32B sets the open-source state of the art at 93.0% on MiniF2F-Test and solves 93 of 672 PutnamBench problems. We release MiniF2F-ALF, an ALF-mutated contamination-sensitive benchmark on which every evaluated model loses accuracy; here our 32B remains strongest and our 4B matches the prior state of the art, Goedel-Prover-V2-32B.

cs.AI↗

Can VLMs Predict Future States? Bootstrapping World Models from Inverse Dynamics

Can unified vision-language models (VLMs) perform forward dynamics prediction (FDP), i.e., predicting the future state (in image form) given the previous observation and an action (in language form)? We find that VLMs struggle to generate physically plausible transitions between frames from instructions. Nevertheless, we identify a crucial asymmetry in multimodal grounding: fine-tuning a VLM to learn inverse dynamics prediction (IDP)-effectively captioning the action between frames-is significantly easier than learning FDP. In turn, IDP can be used to bootstrap FDP through two main strategies: 1) weakly supervised learning from synthetic data and 2) inference time verification. Firstly, IDP can annotate actions for unlabelled pairs of video frame observations to expand the training data scale for FDP. Secondly, IDP can assign rewards to multiple samples of FDP to score them, effectively guiding search at inference time. We evaluate the FDP resulting from both strategies through the task of action-centric image editing on Aurora-Bench with two families of VLMs. Despite remaining general-purpose, our best model achieves a performance competitive with state-of-the-art image editing models, improving on them by a margin between 7% and 13% according to GPT4o-as-judge, and achieving the best average human evaluation across all subsets of Aurora-Bench.

cs.CV↗

Used Car Salesbots? Honesty and Credulity of LLMs as Bargaining Agents under Partial Information

In this work we study agents in simulated bargaining scenarios, where a buyer and a seller communicate through a text channel and attempt to negotiate mutually beneficial trades, under different information regimes (complete information, information asymmetry or mutual uncertainty). We evaluate their performance w.r.t. game-theoretical solutions and further investigate their honesty (their tendency to disclose or withhold information or to mislead and deceive) as well as their credulity (their tendency to trust or distrust information provided by the other agent). We study zero-shot LLM agents with simple prompting scaffolding as well as fine-tuned agents, in order to investigate whether optimising the agents to maximise financial profits makes them stronger negotiators but also more dishonest and less trusting. We find that off-the-shelf LLMs all substantially deviate from game-theoretical equilibria, they attempt to lie about their private information but cannot efficiently exploit information asymmetries. Fine-tuning on financial utility makes the agents stronger at achieving better deals but also more dishonest, highlighting the risks that optimising agents for a task can have on their safety. We release our code and a dataset of bargaining scenarios.

cs.GT↗

Can I Have Your Order? Monte-Carlo Tree Search for Slot Filling Ordering in Diffusion Language Models

While plan-and-infill decoding in Masked Diffusion Models (MDMs) shows promise for mathematical and code reasoning, performance remains highly sensitive to slot infilling order, often yielding substantial output variance. We introduce McDiffuSE, a framework that formulates slot selection as decision making and optimises infilling orders through Monte Carlo Tree Search (MCTS). McDiffuSE uses look-ahead simulations to evaluate partial completions before commitment, systematically exploring the combinatorial space of generation orders. Experiments show an average improvement of 3.2% over autoregressive baselines and 8.0% over baseline plan-and-infill, with notable gains of 19.5% on MBPP and 4.9% on MATH500. Our analysis reveals that while McDiffuSE predominantly follows sequential ordering, incorporating non-sequential generation is essential for maximising performance. We observe that larger exploration constants, rather than increased simulations, are necessary to overcome model confidence biases and discover effective orderings. These findings establish MCTS-based planning as an effective approach for enhancing generation quality in MDMs.

cs.AI↗

Old Habits Die Hard: How Conversational History Geometrically Traps LLMs

How does the conversational past of large language models (LLMs) influence their future performance? Recent work suggests that LLMs are affected by their conversational history in unexpected ways. For instance, hallucinations in prior interactions may influence subsequent model responses. In this work, we introduce History-Echoes, a framework that investigates how conversational history biases subsequent generations. The framework explores this bias from two perspectives: probabilistically, we model conversations as Markov chains to quantify state consistency; geometrically, we measure the consistency of consecutive hidden representations. Across three model families and six datasets spanning diverse phenomena, our analysis reveals a strong correlation between the two perspectives. By bridging these perspectives, we demonstrate that behavioral persistence manifests as a geometric trap, where gaps in the latent space confine the model's trajectory. Code available at https://github.com/technion-cs-nlp/OldHabitsDieHard.

cs.CL↗

Can LLMs Compress (and Decompress)? Evaluating Code Understanding and Execution via Invertibility

LLMs demonstrate strong performance on code benchmarks, yet consistent reasoning across forward and backward execution remains elusive. We present RoundTripCodeEval (RTCE), a benchmark of four code execution reasoning tasks that evaluates round-trip consistency through execution-free, exact-match assessment of bijection fidelity across four lossless compression algorithms. We evaluate state-of-the-art Code-LLMs under zero-shot prompting, supervised fine-tuning on execution traces, and iterative self-reflection. All approaches yield only modest improvements and none closes the gap, revealing that current LLMs lack the internal coherence required for reliable bidirectional code reasoning. RTCE surfaces findings invisible to existing benchmarks: models frequently pass individual forward and backward tasks yet fail the combined round-trip, exposing mutually inconsistent internal representations; SFT and self-reflection saturate after one revision round, indicating they cannot repair fundamental algorithmic misunderstandings; and failures persist even on simple bijections such as RLE, suggesting that algorithmic complexity is not the sole root cause.\footnote{Code and dataset are available at https://github.com/Nickil21/round-trip-code-compression.

cs.LG↗

Lost in Space? Vision-Language Models Struggle with Relative Camera Pose Estimation

We study whether vision-language models (VLMs) can solve relative camera pose estimation (RCPE) from image pairs, a direct test of multi-view spatial reasoning. We cast RCPE as a discrete verbal classification task and introduce \texttt{VRRPI-Bench}, built from real RGB-D frames with object-centric camera motion, and \texttt{VRRPI-Diag}, which isolates individual motion degrees of freedom. Humans (0.91) and specialized geometric pipelines such as LoFTR (0.99) solve the task reliably, yet the best VLM reaches only 0.66 and most others remain near random. Our analyses show that this gap is not basic spatial competence: strong VLMs are near ceiling on single-image benchmarks, but most remain near random once reasoning must span views. They are unstable under source-target reversal (best 59.7\% consistency) and remain weak even in simplified single-DoF settings, especially on optical-axis motions such as roll and depth translation (GPT-5: 0.46 on roll). These failures are useful: they localize concrete missing capabilities, namely cross-view correspondence, view-consistent reasoning, and projective camera-motion understanding, making RCPE a targeted diagnostic for improving multi-view spatial reasoning in VLMs.

cs.CV↗