Searcharxiv⌕ Search

arXiv subjects

Yoshua Bengio

Publications and source records attributed to Yoshua Bengio.

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↗

Towards Trustworthy Physical Intelligence: From Theory to Practice Across Life Cycle

Physical intelligence refers to intelligence systems that understand, reason about and act in accordance with the physical world and its underlying laws, dynamics, and constraints. Unlike conventional AI systems, physical intelligence interacts continuously with uncertain physical environments, and its actions produce consequences that are physically irreversible. As existing trustworthy AI frameworks have been developed primarily for digital AI systems, they do not fully capture the distinctive challenges of Physical Intelligence, such as physical safety, cyber-physical security, and physical manufacturing process. To address this gap, our paper develop a foundation for understanding, operationalizing, and governing trustworthy physical intelligence. First, we characterize the fundamental capabilities of physical intelligence. Second, we investigate the role of physical laws and principles in AI and analyze the associated challenges of physical intelligence. Third, we provide a unified structure of the end-to-end physical intelligence life cycle across five core stages and introduce the Trustworthy Physical Intelligence Operationalization (T-PAIO) framework. Fourth, we formulate the Trustworthy Physical Intelligence (T-PAI) Principles as a theoretical framework that synthesizes key trustworthiness dimensions and establishes a conceptual foundation for the governance of physical intelligence systems.

cs.AI↗

Generative Recursive Reasoning

How should future neural reasoning systems implement extended computation? Recursive Reasoning Models (RRMs) offer a promising alternative to autoregressive sequence extension by performing iterative latent-state refinement with shared transition functions. Yet existing RRMs are largely deterministic, following a single latent trajectory and converging to a single prediction. We introduce Generative Recursive reAsoning Models (GRAM), a framework that turns recursive latent reasoning into probabilistic multi-trajectory computation. GRAM models reasoning as a stochastic latent trajectory, enabling multiple hypotheses, alternative solution strategies, and inference-time scaling through both recursive depth and parallel trajectory sampling. This yields a latent-variable generative model supporting conditional reasoning via $p_θ(y \mid x)$ and, with fixed or absent inputs, unconditional generation via $p_θ(x)$. Trained with amortized variational inference, GRAM improves over deterministic recurrent and recursive baselines on structured reasoning and multi-solution constraint satisfaction tasks, while demonstrating an unconditional generation capability. https://ahn-ml.github.io/gram-website

cs.AI↗

Adaptive Order Policies for Masked Diffusion

Masked diffusion models have seen great success in capturing data distributions over discrete sequences in domains such as text and proteins. These models generate data by iteratively unmasking tokens starting from a fully masked sequence, with the unmasking order typically chosen at random or using a heuristic based on denoiser probabilities. In this work, we propose a scheme for learning the unmasking order using an additional lightweight policy network on top of a diffusion model. Our proposed loss reweights terms in the masked diffusion loss according to policy probabilities, and results in a policy that prefers positions where the denoiser is more likely to be correct. We study this loss in two settings: (i) training solely the policy while using a frozen pre-trained denoiser, and (ii) training the policy and denoiser jointly with the weighted loss to allow for mutual adaptation. We demonstrate that our approach outperforms common heuristics on problems that are sensitive to token ordering, such as combinatorial problems, proteins as well as various coding and language tasks.

cs.LG↗

What if automating AI R&D triggers an intelligence explosion?

In contrast to even a year ago, AI systems now write most of the code inside the companies that build them. As more of the AI research and development (R&D) pipeline is automated, could AI progress radically accelerate in an "intelligence explosion," where years of advances are compressed into months or less? Preliminary evidence suggests that it could. In this work, we assess this evidence, analyze an intelligence explosion's potential impacts, and propose policy responses. AI systems are on track to automate most AI R\&D work within a few years, and possibly all of it. If this triggers an intelligence explosion, it could dramatically bring forward AI's benefits, but also pose extreme risks: capabilities growth could accelerate far beyond what society can keep up with, humanity could lose control over superhuman AI systems, and checks on power within and between states, companies, and branches of government could be severely eroded. Although there remains much uncertainty about these possibilities, the high stakes warrant serious further attention. Policymakers should urgently obtain more visibility into the automation of AI R&D, develop ways to steer and constrain an intelligence explosion, and prepare society to adapt to an intelligence explosion's impacts.

cs.CY↗

Distribution-Based Regularization for Reward Fine-Tuning Diffusion and Flow Models

Reward fine-tuning is an effective way to adapt pretrained diffusion and flow-matching models to downstream objectives, but aggressive optimization often leads to mode collapse and distributional drift. Existing methods attempt to control this drift with reverse-KL regularization, implemented as an $\ell_2$ penalty on the velocity field; however, this functions by indirectly constraining sampling dynamics rather than directly acting on the generated distribution. To address this, we propose DFR (Distributional Functional Rewards), which augments the task reward with a distribution-dependent, per-sample reward that acts explicitly on the generated distribution. Formally, we generalize the KL-regularized objective from expected rewards, which are linear in the model's distribution, to convex functionals of it. The first variation of this objective defines an effective, distribution-dependent reward that directly penalizes deviation from the pretrained distribution. This formulation yields additive per-sample rewards for the Fréchet distance (FD), maximum mean discrepancy (MMD), and DreamSim diversity, recovering several recently proposed rewards as special cases. Across various text-to-image fine-tuning tasks, DFR matches the peak reward of KL-tuned baselines while increasing diversity by up to 29% at equal or lower drift.

cs.LG↗

Evaluating Predictor-Corrector Dependence in Hard-Constrained Climate Emulators

Climate emulators often enforce physical constraints by correcting each state predicted by a neural network, but evaluating only the corrected forecast conflates predictor skill with the effect of the correction. We introduce an evaluation framework that separates these contributions by comparing corrected forecasts, pre-correction predictions along corrected trajectories, and autonomous rollouts with correction disabled. We further cross correction during training and inference to distinguish its immediate effect from how it shapes the learned predictor, and use matched nonphysical controls to test whether improvements are specific to the physical correction. In ACE2, correction substantially improves physical consistency. In a three-seed ACE2 pilot, models trained with correction produce less accurate uncorrected precipitation forecasts at short horizons than models trained without correction, but more accurate forecasts at 10 and 30 days. This contradicts our preregistered hypothesis that the deficit would grow with forecast horizon. In a separate nine-seed experiment, directly supervising the uncorrected output reverses its precipitation forecast deficit at 72 hours while maintaining the accuracy of corrected forecasts at 24 hours. In ConCerNet's heat-equation system, projection-trained models have higher uncorrected trajectory-based vector-field RMSE than models trained without projection, while restoring the physical projection recovers accuracy far more effectively than a matched nonphysical control. Together, these results show that hard corrections can reshape the predictors trained alongside them, creating dependencies that are not visible from corrected forecasts alone. Hard-constrained emulators should therefore be evaluated by separately measuring predictor accuracy, correction effects, and their interaction during training.

cs.CY↗

GCAM-Gym: A Benchmark for Agent-Based Climate Policy Search

Comparing climate-policy search methods requires accounting for differences in policy control, available observations, feedback timing, and simulator budgets. We introduce GCAM-Gym, a Gymnasium-compatible environment that supports full-policy evaluation and stepwise control through a common interface and exposes the components of the policy objective. We compare seven search and control systems in three settings: target attainment and cost, cost-temperature trade-offs, and one-step delayed feedback. Each system is evaluated with ten seeds per setting, for a total of 3,210 attempted GCAM evaluations. Four systems share a budget of 24 attempts per seed; among these, CMA-ES achieves the highest interquartile mean (IQM) of the best returns from completed policy evaluations in all three settings. We release the evaluation archive and analysis tools to reconstruct the reported comparisons without running GCAM.

cs.CY↗

Language models recognize dropout and Gaussian noise applied to their activations

We provide evidence that language models can detect, localize and, to a degree, verbalize the difference between a-semantic perturbations applied to their activations. More precisely, we either (a) apply dropout or (b) add Gaussian noise to them, at a target sentence. We then ask a multiple-choice question such as "Which of the previous sentences was perturbed?" or "Which perturbation was applied?". We test eight models from the Gemma, Llama, Mistral, Olmo, Qwen and Seed families, with sizes between 8B and 36B, all of which can easily detect and localize the perturbations, often with perfect accuracy. To investigate the mechanisms, we find that Olmo models develop this capacity during pre-training, and we rule out the hypothesis that perturbed sentences simply have lower log-likelihood. Notably, the zero-shot accuracy of Qwen3-32B and Qwen3.8-27B in identifying which perturbation was applied correlates with the perturbation strength. While this task proves hard for the remaining models, most of them can learn, when taught in context, to distinguish between dropout and Gaussian noise. Interestingly, Qwen3-32B and Qwen3.8-27B learn worse when supervised with wrong labels, again suggesting a prior towards the correct ones. Because dropout is used as a training-regularization technique, while Gaussian noise is sometimes added during inference, we discuss the possibility of a data-agnostic "training awareness" signal and the implications for AI safety.

cs.AI↗

Defining and Generating Compositional Structure

Compositional structure is pervasive in nature and across scales: atoms combine into molecules, molecules into cells, and cells into organisms, while further compositional hierarchies appear within each level. More broadly, compositionality, the reuse and recombination of parts into novel structures, is a fundamental source of interesting complexity in both natural and artificial systems. Yet despite its importance, and despite our apparent ease in recognizing it, mathematical accounts capturing what we intuitively mean by compositional structure remain limited. Here, we fill this gap by formalizing a simple idea: compositional structure emerges when parts, considered together, reveal shared structure that is not captured when those parts are modeled separately. We express this idea using algorithmic information theory, the mathematics of information and compression, which provides a natural way to identify structure in arbitrary data without specifying in advance what that structure must look like. Applied across scales, our definition informs where and how compositional structure emerges within an object. Using tractable approximations, we then invert the framework to generate compositional objects: we specify the desired form of their compositional structure without prescribing their particular content, and optimize directly over strings and images. The resulting objects exhibit readily recognizable hierarchical reuse and part-whole organization, suggesting that the definition captures compositional structure rather than merely complexity, regularity, or shared information. Finally, we discuss implications for machine learning, including a new perspective on depth in neural architectures and principles for constructing compositional data curricula capable of supporting open-ended learning.

cs.AI↗

The Smart Buildings Challenge: Learning Foundation Models for Real-World Optimization at Scale

Buildings cause nearly 40\% of carbon emissions in the US. Of total building emissions, HVAC systems are responsible for approximately 40-60\% and account for 15\% of global energy consumption. Despite advances in artificial intelligence, most buildings still operate as inefficiently as they did decades ago. Building optimization, perhaps more so than any other optimization problem, has enormous potential for Machine Learning (ML) driven impact on climate change, in both the short and long term. One of the biggest obstacles for building optimization is the lack of policy transfer. Most building control systems, whether simulation or learning based, don't generalize: they only work well for the building they were developed for, but not for other buildings. To solve this problem, we propose a challenge for the development of building foundation models that easily adapt to new buildings with minimal configuration and thereby enable building optimization to scale across the globe. Participants will build control policies on a set of simulated buildings, and have their policies evaluated on an unseen set of buildings. The buildings vary in size, layout, weather, material properties, and HVAC system. Successful outcomes from this initiative promise a massive leap forward in global sustainability efforts and demonstrate the great potential of ML to combat climate change.

cs.AI↗

Multi-Objective Amortized CDRH3 Sequence Sampling Diversifies Promising Antibody Candidates

Antibody therapeutics have transformed treatment across diverse therapeutic areas and have become central to modern medicine. While recent machine learning approaches aim to accelerate antibody discovery through structure-based design and target-affinity optimization, successful candidates must also satisfy broad developability constraints, including structural compatibility, low immunogenicity, and favorable biophysical properties, while remaining sufficiently diverse to provide sufficient coverage of design space to mitigate later-stage development failures. We introduce a preference-conditioned multi-objective generative flow network (GFN) fine-tuning pipeline for fixed-framework CDRH3 loop redesign that fine-tunes a biological sequence model against many inexpensive proxies covering backbone compatibility, sequence likelihood, immunogenicity risk, novelty, and developability liabilities. It combines up to fourteen predictive rewards in a single policy trained to sample CDRH3 loops across 623 SAbDab antibody structures, leveraging a reward reduction strategy to preserve Pareto dominance under this learned map. We also introduce a novel autoregressive span-in-filling variant of AMPLIFY-350M that serves both as a scaffold-conditioned CDRH3 generation prior and as a direct conditional likelihood scorer. We evaluated reward scores, diversity, and design quality on loops sampled on 45 held-out antibody complexes to assess structural consistency, confidence, aggregation propensity, stability, and antigen binding. Compared with the inverse-folding methods ProteinMPNN and ESM-IF, our approach produces roughly 10x more distinct acceptable candidates per design at 50% sequence identity, while holding 10% of the 14-objective Pareto front pooled across methods, trading per-sample quality for greater exploration of novel possible structures.

cs.CE↗

Mixture of GFlowNets

Generative Flow Networks (GFlowNets) learn a policy that samples compositional objects in proportion to an unnormalized reward $R(x)$. This makes them useful wherever many diverse, high-reward candidates are needed and oracle queries are expensive. In practice, a single GFlowNet often fails to cover all modes of a complex distribution, and the obvious ways to scale compute saturate: larger batches and data-parallel gradient averaging draw every trajectory from one, possibly biased, policy, so extra rollouts rediscover known modes instead of finding new ones. We explore whether this mode-bias problem (mode collapse in the extreme) can be addressed with a population of asynchronously trained policies which divide the state space between them. Mixture of GFlowNets (MoG) replaces the single forward policy with a community of $K$ independent GFlowNets coupled through one jointly trained soft classifier, which assigns each member its own region of the state space through reward shaping. Mixing the members under their learned partition functions provably recovers the target $p(x)\propto R(x)$. MoG exchange (MoGX) also lets the classifier route sampled trajectories: a trajectory the classifier assigns to another member is sent to that member and trained on there, with no importance sampling. We evaluate both methods by keeping the environment-interaction budget fixed while scaling the community size, turning parallel compute into mode coverage without extra oracle calls. On HyperGrid, across six reward landscapes and communities of up to $K{=}64$, MoG and MoGX keep finding new modes as $K$ grows, while data-parallel training saturates. On a genomic ideotype-design task, MoGX covers $39%$ of the $1024$ optima at $K{=}64$, against $31\%$ for independent agents and $20\%$ for data-parallel training. MoG also has a lower per-iteration synchronization cost than data-parallel training.

cs.LG↗

Scaling Latent Reasoning via Looped Language Models

Modern LLMs are trained to "think" primarily via explicit text generation, such as chain-of-thought (CoT), which defers reasoning to post-training and under-leverages pre-training data. We present and open-source Ouro, named after the recursive Ouroboros, a family of pre-trained Looped Language Models (LoopLM) that instead build reasoning into the pre-training phase through (i) iterative computation in latent space, (ii) an entropy-regularized objective for learned depth allocation, and (iii) scaling to 7.7T tokens. Ouro 1.4B and 2.6B models enjoy superior performance that match the results of up to 12B SOTA LLMs across a wide range of benchmarks. Through controlled experiments, we show this advantage stems not from increased knowledge capacity, but from superior knowledge manipulation capabilities. We also show that LoopLM yields reasoning traces more aligned with final outputs than explicit CoT. We hope our results show the potential of LoopLM as a novel scaling direction in the reasoning era. Our model is available here: http://ouro-llm.github.io.

cs.CL↗

PyINE: A Framework for Scalable Elicitation and Oversight via Code Execution

Reasoning models can remain capable of solving a task while still defaulting to cheaper but misleading shortcuts. This creates a central oversight problem: when a model gives an answer with plausible but incomplete reasoning, can an overseer determine whether that output should be trusted? To study this problem, we introduce PyINE, a framework for scalable elicitation and oversight using instrumented Python programs as a verifiable execution substrate. In PyINE, programs define task environments, execution traces provide authoritative labels for outcomes and intermediate facts, and task variants can be generated mechanically rather than through static human annotation. We instantiate the framework in PyINE-v1, a first release built from nearly one million deterministic execution traces and over 500,000 matched LLM-generated code variants used for counterfactual evaluation. Using standard RL with verifiable rewards on cue-varied tasks, we train a shortcut-following model that improves substantially at predicting execution outcomes while still making systematic errors when misleading human-facing cues conflict with the program's realized behavior. We then evaluate activation probes, trained text classifiers, prompted judges, and a lightweight debate protocol as overseers of this model. We find that performance pooled at the dataset level can hide weak coverage of the failures that matter most: cheap learned overseers often miss rare shortcut-driven errors, while stronger model-based checks are more balanced but substantially costlier and harder to turn into reliable thresholded decisions. PyINE-v1 turns this failure-mode coverage problem into a reusable experimental setting for developing oversight methods that are verifiable, failure-mode-aware, and cost-sensitive.

cs.AI↗

Machine learning and information theory concepts towards an AI Mathematician

The current state-of-the-art in artificial intelligence is impressive, especially in terms of mastery of language, but not so much in terms of mathematical reasoning. What could be missing? Can we learn something useful about that gap from how the brains of mathematicians go about their craft? This essay builds on the idea that current deep learning mostly succeeds at system 1 abilities -- which correspond to our intuition and habitual behaviors -- but still lacks something important regarding system 2 abilities -- which include reasoning and robust uncertainty estimation. It takes an information-theoretical posture to ask questions about what constitutes an interesting mathematical statement, which could guide future work in crafting an AI mathematician. The focus is not on proving a given theorem but on discovering new and interesting conjectures. The central hypothesis is that a desirable body of theorems better summarizes the set of all provable statements, for example by having a small description length while at the same time being close (in terms of number of derivation steps) to many provable statements.

cs.AI↗

CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs

AI control protocols oversee untrusted models by monitoring their actions and modifying potentially unsafe steps, often using a trusted model. This partially tampers with the untrusted model's trajectory. If the acting model detects such an intervention, it may infer properties of the monitor and adapt to evade the control protocol. We introduce CIAware-Bench, a benchmark for measuring control intervention (CI) awareness across frontier models. CIAware-Bench tests whether models can distinguish their own trajectories from those modified by a control intervention. The benchmark comprises four task domains (essay writing, BigCodeBench, Bash Arena, and SHADE-Arena), with options to vary trajectory watermarking, side-task presence, and the control protocol. Evaluating thirteen acting models with eight intervening models shows substantial variation between settings and model pairs. CI awareness rises sharply for GPT-6 Astra and the Claude 5 models (Fable 5 and Opus 5). When explicitly probed, Astra reaches mean AUROC of 0.90 on Essay, 0.91 on BigCodeBench, 0.86 on Bash Arena and 0.77 on SHADE-Arena. Fable 5 and Opus 5 both reach 0.77 on Essay, with less consistent gains in the other environments. On SHADE-Arena, we demonstrate that telling an acting model that an action was replaced and asking it to adapt leads to moderate improvements in monitor evasion rates. In summary, our results highlight that control evaluations should already assume perfect CI awareness for conservative safety estimates, and that protocol design should explore countermeasures that make interventions harder to detect.

cs.AI↗

AI Safety: Not Optional, Not Later

Incidents show that AI safety failures often arise across multiple layers. We present a safety-by-design assurance architecture combining model-level supervision, such as Scientist AI, with system-level controls over scaffolds and harnesses, independent verification, monitoring, and evidence infrastructure, supported by governance for accountability and evidence interoperability.

cs.SE↗