arXiv subjects

Yarin Gal

Publications and source records attributed to Yarin Gal.

At least 37 records · Page 2Linked to original sources

arXiv2026

Stabilizing Policy Gradients for Sample-Efficient Reinforcement Learning in LLM Reasoning

Reinforcement Learning, particularly through policy gradient methods, has played a central role in enabling reasoning capabilities of Large Language Models. However, the optimization stability of policy gradients in this setting remains understudied. As a result, existing implementations often resort to conservative hyperparameter choices to ensure stability, which requires more training samples and increases computational costs. Hence, developing models for reliably tracking the underlying optimization dynamics and leveraging them into training enables more sample-efficient regimes and further unleashes scalable post-training. We address this gap by formalizing the stochastic optimization problem of policy gradients with explicit consideration of second-order geometry. We propose a tractable computational framework that tracks and leverages curvature information during policy updates. We further employ this framework to design interventions in the optimization process through data selection. The resultant algorithm, Curvature-Aware Policy Optimization (CAPO), identifies samples that contribute to unstable updates and masks them out. Theoretically, we establish monotonic improvement guarantees under realistic assumptions. On standard math reasoning benchmarks, we empirically show that CAPO ensures stable updates under aggressive learning regimes where baselines catastrophically fail. With minimal intervention (rejecting fewer than 8% of tokens), CAPO achieves up to 30x improvement in sample efficiency over standard GRPO for LLM reasoning.

cs.LG↗

arXiv2026

OMNI-LEAK: Orchestrator Multi-Agent Network Induced Data Leakage

As Large Language Model (LLM) agents become more capable, their coordinated use in the form of multi-agent systems is anticipated to emerge as a practical paradigm. Prior work has examined the safety and misuse risks associated with agents. However, much of this has focused on the single-agent case and/or setups missing basic engineering safeguards such as access control, revealing a scarcity of threat modeling in multi-agent systems. We investigate the security vulnerabilities of a popular multi-agent pattern known as the orchestrator setup, in which a central agent decomposes and delegates tasks to specialized agents. Through red-teaming a concrete setup representative of a likely future use case, we demonstrate a novel attack vector, OMNI-LEAK, that compromises several agents to leak sensitive data through a single indirect prompt injection, even in the presence of data access control. We report the susceptibility of frontier models to different categories of attacks, finding that both reasoning and non-reasoning models are vulnerable, even when the attacker lacks insider knowledge of the implementation details. Our work highlights the importance of safety research to generalize from single-agent to multi-agent settings, in order to reduce the serious risks of real-world privacy breaches and financial losses and overall public trust in AI agents.

cs.AI↗

arXiv2026

Boundary Point Jailbreaking of Black-Box LLMs

Frontier LLMs are safeguarded against attempts to extract harmful information via adversarial prompts known as "jailbreaks". Recently, defenders have developed classifier-based systems that have survived thousands of hours of human red teaming. We introduce Boundary Point Jailbreaking (BPJ), a new class of automated jailbreak attacks that evade the strongest industry-deployed safeguards. Unlike previous attacks that rely on white/grey-box assumptions (such as classifier scores or gradients) or libraries of existing jailbreaks, BPJ is fully black-box and uses only a single bit of information per query: whether or not the classifier flags the interaction. To achieve this, BPJ addresses the core difficulty in optimising attacks against robust real-world defences: evaluating whether a proposed modification to an attack is an improvement. Instead of directly trying to learn an attack for a target harmful string, BPJ converts the string into a curriculum of intermediate attack targets and then actively selects evaluation points that best detect small changes in attack strength ("boundary points"). We believe BPJ is the first fully automated attack algorithm that succeeds in developing universal jailbreaks against Constitutional Classifiers, as well as the first automated attack algorithm that succeeds against GPT-5's input classifier without relying on human attack seeds. BPJ is difficult to defend against in individual interactions but incurs many flags during optimisation, suggesting that effective defence requires supplementing single-interaction methods with batch-level monitoring.

cs.LG↗

arXiv2026

Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs

Open-weight AI systems offer unique benefits, including enhanced transparency, open research, and decentralized access. However, they are vulnerable to tampering attacks which can efficiently elicit harmful behaviors by modifying weights or activations. Currently, there is not yet a robust science of open-weight model risk management. Existing safety fine-tuning methods and other post-training techniques have struggled to make LLMs resistant to more than a few dozen steps of adversarial fine-tuning. In this paper, we investigate whether filtering text about dual-use topics from training data can prevent unwanted capabilities and serve as a more tamper-resistant safeguard. We introduce a multi-stage pipeline for scalable data filtering and show that it offers a tractable and effective method for minimizing biothreat proxy knowledge in LLMs. We pretrain multiple 6.9B-parameter models from scratch and find that they exhibit substantial resistance to adversarial fine-tuning attacks on up to 10,000 steps and 300M tokens of biothreat-related text -- outperforming existing post-training baselines by over an order of magnitude -- with no observed degradation to unrelated capabilities. However, while filtered models lack internalized dangerous knowledge, we find that they can still leverage such information when it is provided in context (e.g., via search tool augmentation), demonstrating a need for a defense-in-depth approach. Overall, these findings help to establish pretraining data curation as a promising layer of defense for open-weight AI systems.

cs.LG↗

arXiv2026

MIP against Agent: Malicious Image Patches Hijacking Multimodal OS Agents

Recent advances in operating system (OS) agents have enabled vision-language models (VLMs) to directly control a user's computer. Unlike conventional VLMs that passively output text, OS agents autonomously perform computer-based tasks in response to a single user prompt. OS agents do so by capturing, parsing, and analysing screenshots and executing low-level actions via application programming interfaces (APIs), such as mouse clicks and keyboard inputs. This direct interaction with the OS significantly raises the stakes, as failures or manipulations can have immediate and tangible consequences. In this work, we uncover a novel attack vector against these OS agents: Malicious Image Patches (MIPs), adversarially perturbed screen regions that, when captured by an OS agent, induce it to perform harmful actions by exploiting specific APIs. For instance, a MIP can be embedded in a desktop wallpaper or shared on social media to cause an OS agent to exfiltrate sensitive user data. We show that MIPs generalise across user prompts and screen configurations, and that they can hijack multiple OS agents even during the execution of benign instructions. These findings expose critical security vulnerabilities in OS agents that have to be carefully addressed before their widespread deployment.

cs.CR↗

arXiv2026

Weighted Conditional Flow Matching

Conditional flow matching (CFM) has emerged as a powerful framework for training continuous normalizing flows due to its computational efficiency and effectiveness. However, standard CFM often produces paths that deviate significantly from straight-line interpolations between prior and target distributions, making generation slower and less accurate due to the need for fine discretization at inference. Recent methods enhance CFM performance by inducing shorter and straighter trajectories but typically rely on computationally expensive mini-batch optimal transport (OT). Drawing insights from entropic optimal transport (EOT), we propose Weighted Conditional Flow Matching (W-CFM), a novel approach that modifies the classical CFM loss by weighting each training pair $(x, y)$ with a Gibbs kernel. We show that this weighting recovers the entropic OT coupling up to some bias in the marginals, and we provide the conditions under which the marginals remain nearly unchanged. Moreover, we establish an equivalence between W-CFM and the minibatch OT method in the large-batch limit, showing how our method overcomes computational and performance bottlenecks linked to batch size. Empirically, we test our method on unconditional generation on various synthetic and real datasets, confirming that W-CFM achieves comparable or superior sample quality, fidelity, and diversity to other alternative baselines while maintaining the computational efficiency of vanilla CFM.

cs.LG↗

arXiv2025

Iterative Deployment Improves Planning Skills in LLMs

We show that iterative deployment of large language models (LLMs), each fine-tuned on data carefully curated by users from the previous models' deployment, can significantly change the properties of the resultant models. By testing this mechanism on various planning domains, we observe substantial improvements in planning skills, with later models displaying emergent generalization by discovering much longer plans than the initial models. We then provide theoretical analysis showing that iterative deployment effectively implements reinforcement learning (RL) training in the outer-loop (i.e. not as part of intentional model training), with an implicit reward function. The connection to RL has two important implications: first, for the field of AI safety, as the reward function entailed by repeated deployment is not defined explicitly, and could have unexpected implications to the properties of future model deployments. Second, the mechanism highlighted here can be viewed as an alternative training regime to explicit RL, relying on data curation rather than explicit rewards.

cs.AI↗

arXiv2025

A Definition of AGI

The lack of a concrete definition for Artificial General Intelligence (AGI) obscures the gap between today's specialized AI and human-level cognition. This paper introduces a quantifiable framework to address this, defining AGI as matching the cognitive versatility and proficiency of a well-educated adult. To operationalize this, we ground our methodology in Cattell-Horn-Carroll theory, the most empirically validated model of human cognition. The framework dissects general intelligence into ten core cognitive domains-including reasoning, memory, and perception-and adapts established human psychometric batteries to evaluate AI systems. Application of this framework reveals a highly "jagged" cognitive profile in contemporary models. While proficient in knowledge-intensive domains, current AI systems have critical deficits in foundational cognitive machinery, particularly long-term memory storage. The resulting AGI scores (e.g., GPT-4 at 27%, GPT-5 at 57%) concretely quantify both rapid progress and the substantial gap remaining before AGI.

cs.AI↗

arXiv2025

Memo: Training Memory-Efficient Embodied Agents with Reinforcement Learning

To enable embodied agents to operate effectively over extended timeframes, it is crucial to develop models that form and access memories to stay contextualized in their environment. In the current paradigm of training transformer-based policies for embodied sequential decision-making tasks, visual inputs often overwhelm the context limits of transformers, while humans can maintain and utilize a lifetime of experience compressed as memories. Significant compression is possible in principle, as much of the input is irrelevant and can be abstracted. However, existing approaches predominantly focus on either recurrent models with fixed-size memory or transformers with full-context reliance. In this work, we propose Memo, a transformer-based architecture and training recipe for reinforcement learning (RL) on memory-intensive, long-horizon tasks. Memo incorporates the creation and retrieval of memory by interleaving periodic summarization tokens with the inputs of a model during training. We demonstrate Memo's effectiveness on a gridworld meta-RL benchmark and a multi-object navigation task in photo-realistic indoor settings. Memo outperforms naive long-context transformer baselines while being more compute and storage efficient. Additionally, Memo generalizes better to longer contexts at inference time and remains robust in streaming settings, where historical context must be truncated to fit inference constraints. Our code is available at: https://github.com/gunshi/memo.

cs.AI↗

arXiv2025

Scaling Up Active Testing to Large Language Models

Active testing enables label-efficient evaluation of predictive models through careful data acquisition, but it can pose a significant computational cost. We identify cost-saving measures that enable active testing to be scaled up to large language models (LLMs). In particular we show that the surrogate model used to guide data acquisition can be constructed cheaply using in-context learning, does not require updating within an active-testing loop, and can be smaller than the target model. We even find we can make good data-acquisition decisions without making predictions with the target model. As a result we are able to achieve much more accurate evaluations of LLM performance relative to using randomly acquired data. We additionally introduce a bootstrap estimator of evaluation error, which we show to be a useful indicator of how well active testing is working within a single run.

cs.LG↗

arXiv2025

RedCodeAgent: Automatic Red-teaming Agent against Diverse Code Agents

Code agents have gained widespread adoption due to their strong code generation capabilities and integration with code interpreters, enabling dynamic execution, debugging, and interactive programming capabilities. While these advancements have streamlined complex workflows, they have also introduced critical safety and security risks. Current static safety benchmarks and red-teaming tools are inadequate for identifying emerging real-world risky scenarios, as they fail to cover certain boundary conditions, such as the combined effects of different jailbreak tools. In this work, we propose RedCodeAgent, the first automated red-teaming agent designed to systematically uncover vulnerabilities in diverse code agents. With an adaptive memory module, RedCodeAgent can leverage existing jailbreak knowledge, dynamically select the most effective red-teaming tools and tool combinations in a tailored toolbox for a given input query, thus identifying vulnerabilities that might otherwise be overlooked. For reliable evaluation, we develop simulated sandbox environments to additionally evaluate the execution results of code agents, mitigating potential biases of LLM-based judges that only rely on static code. Through extensive evaluations across multiple state-of-the-art code agents, diverse risky scenarios, and various programming languages, RedCodeAgent consistently outperforms existing red-teaming methods, achieving higher attack success rates and lower rejection rates with high efficiency. We further validate RedCodeAgent on real-world code assistants, e.g., Cursor and Codeium, exposing previously unidentified security risks. By automating and optimizing red-teaming processes, RedCodeAgent enables scalable, adaptive, and effective safety assessments of code agents.

cs.SE↗

arXiv2025

Measuring what Matters: Construct Validity in Large Language Model Benchmarks

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as 'safety' and 'robustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

cs.CL↗

arXiv2025

Fundamental Limitations in Pointwise Defences of LLM Finetuning APIs

LLM developers have imposed technical interventions to prevent fine-tuning misuse attacks, attacks where adversaries evade safeguards by fine-tuning the model using a public API. Previous work has established several successful attacks against specific fine-tuning API defences. In this work, we show that defences of fine-tuning APIs that seek to detect individual harmful training or inference samples ('pointwise' detection) are fundamentally limited in their ability to prevent fine-tuning attacks. We construct 'pointwise-undetectable' attacks that repurpose entropy in benign model outputs (e.g. semantic or syntactic variations) to covertly transmit dangerous knowledge. Our attacks are composed solely of unsuspicious benign samples that can be collected from the model before fine-tuning, meaning training and inference samples are all individually benign and low-perplexity. We test our attacks against the OpenAI fine-tuning API, finding they succeed in eliciting answers to harmful multiple-choice questions, and that they evade an enhanced monitoring system we design that successfully detects other fine-tuning attacks. We encourage the community to develop defences that tackle the fundamental limitations we uncover in pointwise fine-tuning API defences.

cs.LG↗

arXiv2025

Temporal-Difference Variational Continual Learning

Machine Learning models in real-world applications must continuously learn new tasks to adapt to shifts in the data-generating distribution. Yet, for Continual Learning (CL), models often struggle to balance learning new tasks (plasticity) with retaining previous knowledge (memory stability). Consequently, they are susceptible to Catastrophic Forgetting, which degrades performance and undermines the reliability of deployed systems. In the Bayesian CL literature, variational methods tackle this challenge by employing a learning objective that recursively updates the posterior distribution while constraining it to stay close to its previous estimate. Nonetheless, we argue that these methods may be ineffective due to compounding approximation errors over successive recursions. To mitigate this, we propose new learning objectives that integrate the regularization effects of multiple previous posterior estimations, preventing individual errors from dominating future posterior updates and compounding over time. We reveal insightful connections between these objectives and Temporal-Difference methods, a popular learning mechanism in Reinforcement Learning and Neuroscience. Experiments on challenging CL benchmarks show that our approach effectively mitigates Catastrophic Forgetting, outperforming strong Variational CL methods.

cs.LG↗

arXiv2025

Evaluating & Reducing Deceptive Dialogue From Language Models with Multi-turn RL

Large Language Models (LLMs) interact with millions of people worldwide in applications such as customer support, education and healthcare. However, their ability to produce deceptive outputs, whether intentionally or inadvertently, poses significant safety concerns. The unpredictable nature of LLM behavior, combined with insufficient safeguards against hallucination, misinformation, and user manipulation, makes their misuse a serious, real-world risk. In this paper, we investigate the extent to which LLMs engage in deception within dialogue, and propose the belief misalignment metric to quantify deception. We evaluate deception across four distinct dialogue scenarios, using five established deception detection metrics and our proposed metric. Our findings reveal this novel deception measure correlates more closely with human judgments than any existing metrics we test. Additionally, our benchmarking of eight state-of-the-art models indicates that LLMs naturally exhibit deceptive behavior in approximately 26% of dialogue turns, even when prompted with seemingly benign objectives. When prompted to deceive, LLMs are capable of increasing deceptiveness by as much as 31% relative to baselines. Unexpectedly, models trained with RLHF, the predominant approach for ensuring the safety of widely-deployed LLMs, still exhibit deception at a rate of 43% on average. Given that deception in dialogue is a behavior that develops over an interaction history, its effective evaluation and mitigation necessitates moving beyond single-utterance analyses. We introduce a multi-turn reinforcement learning methodology to fine-tune LLMs to reduce deceptive behaviors, leading to a 77.6% reduction compared to other instruction-tuned models.

cs.CL↗

arXiv2025

Estimating the Hallucination Rate of Generative AI

This paper presents a method for estimating the hallucination rate for in-context learning (ICL) with generative AI. In ICL, a conditional generative model (CGM) is prompted with a dataset and a prediction question and asked to generate a response. One interpretation of ICL assumes that the CGM computes the posterior predictive of an unknown Bayesian model, which implicitly defines a joint distribution over observable datasets and latent mechanisms. This joint distribution factorizes into two components: the model prior over mechanisms and the model likelihood of datasets given a mechanism. With this perspective, we define a hallucination as a generated response to the prediction question with low model likelihood given the mechanism. We develop a new method that takes an ICL problem and estimates the probability that a CGM will generate a hallucination. Our method only requires generating prediction questions and responses from the CGM and evaluating its response log probability. We empirically evaluate our method using large language models for synthetic regression and natural language ICL tasks.

cs.LG↗

arXiv2025

Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples

Poisoning attacks can compromise the safety of large language models (LLMs) by injecting malicious documents into their training data. Existing work has studied pretraining poisoning assuming adversaries control a percentage of the training corpus. However, for large models, even small percentages translate to impractically large amounts of data. This work demonstrates for the first time that poisoning attacks instead require a near-constant number of documents regardless of dataset size. We conduct the largest pretraining poisoning experiments to date, pretraining models from 600M to 13B parameters on chinchilla-optimal datasets (6B to 260B tokens). We find that 250 poisoned documents similarly compromise models across all model and dataset sizes, despite the largest models training on more than 20 times more clean data. We also run smaller-scale experiments to ablate factors that could influence attack success, including broader ratios of poisoned to clean data and non-random distributions of poisoned samples. Finally, we demonstrate the same dynamics for poisoning during fine-tuning. Altogether, our results suggest that injecting backdoors through data poisoning may be easier for large models than previously believed as the number of poisons required does not scale up with model size, highlighting the need for more research on defences to mitigate this risk in future models.

cs.LG↗

arXiv2025

Detecting LLM Hallucination Through Layer-wise Information Deficiency: Analysis of Ambiguous Prompts and Unanswerable Questions

Large language models (LLMs) frequently generate confident yet inaccurate responses, introducing significant risks for deployment in safety-critical domains. We present a novel, test-time approach to detecting model hallucination through systematic analysis of information flow across model layers. We target cases when LLMs process inputs with ambiguous or insufficient context. Our investigation reveals that hallucination manifests as usable information deficiencies in inter-layer transmissions. While existing approaches primarily focus on final-layer output analysis, we demonstrate that tracking cross-layer information dynamics ($\mathcal{L}$I) provides robust indicators of model reliability, accounting for both information gain and loss during computation. $\mathcal{L}$I integrates easily with pretrained LLMs without requiring additional training or architectural modifications.

cs.LG↗