arXiv subjects

Yonathan Efroni

Publications and source records attributed to Yonathan Efroni.

At least 19 recordsLinked to original sources

arXiv2026

Towards Scaling Reinforcement Learning to Massive Populations: Learning Mean-Field Representations

Modern multi-agent systems are increasingly deployed at scale over large populations of agents in settings such as ad-auctions, traffic routing, and recommendation systems. The dominant approach in such settings is to optimize each agent's policy independently, treating the other agents as part of a fixed single-agent environment rather than modeling the population dynamics. In many large-population systems, the dynamics depend on an aggregate summary of the population rather than the identity of any individual. Mean-field RL exploits such structure, providing a principled framework that models each agent's environment as an explicit function of the population distribution. However, in large state-action spaces or high-dimensional control problems, modeling the population distribution is itself intractable. How can we design a scalable framework for high-dimensional control problems with large populations? This work explores this question from the perspective of representation learning. We introduce a mean-field RL framework in which the rewards and transition dynamics depend on the population only through an unknown low-dimensional aggregate statistic. We then study this framework in the offline setting and design a provable approach that learns a near-optimal policy by learning a low-dimensional representation. Motivated by real-life supply-chain optimization problems, we design a one-step routing game to test the hypothesis that learning a low-dimensional population representation improves reward prediction and Nash gap estimation relative to baselines that don't exploit this structure. We show that under a fixed neural-network parameter count and optimization budget, learning a low-dimensional population representation improves reward prediction and the equilibrium quality of the resulting policies.

cs.MA↗

arXiv2026

Hack-Verifiable Terminal Bench: Evaluating Reward Hacking in Terminal Tasks

As agents grow more capable and autonomous, their tendency to reward hack, satisfying a task's checks while violating its intent, becomes an increasingly important failure mode. Measuring reward hacking is itself challenging, as detection typically relies on human inspection or LLM judges, both of which can be unreliable. The hack-verifiable environments (HVE) methodology addresses this challenge by embedding detectable hacks into tasks, allowing reward hacks to be identified automatically and reliably. In this work, we adapt HVE to Terminal Bench, a leading benchmark of real-world terminal and coding tasks, and introduce Hack-Verifiable Terminal Bench (HVTB). Using HVTB, we measure reward-hacking rates across frontier models and study whether prompts with varying amounts of information on the hack can mitigate this behavior. This lets us test whether prompting can prevent not only known reward-hacking strategies, but also 'unknown unknown' exploits that the prompt does not anticipate. We release all environments and agent traces at https://majoroth.github.io/hack-verifiable-environments/hvtb

cs.AI↗

arXiv2026

Frustratingly Simple Black-Box Adaptation of Language Models via Logit Bias

Many organizations aim to adapt language models for internal use, both to improve performance on domain-specific tasks and to address privacy concerns around sensitive data. However, such adaptation remains non-trivial: it often requires operationally challenging fine-tuning of open-source models or ad hoc prompt optimization. We study a minimal alternative based on a simple API-level control: allowing users to bias the model's logits with a user-defined vector. We develop a black-box method for learning a single context-independent logit-bias vector, added at every decoding step, without modifying model weights or requiring gradients. Starting from a KL-regularized reinforcement learning (RL) objective, we characterize when such a fixed logit-bias vector can approximate the optimal prefix-dependent correction and derive a closed-form inverse-propensity estimator from rollouts, rewards, and token probabilities. Empirically, this simple decoding-time intervention improves over base models on mathematical and reasoning benchmarks while using far fewer trainable parameters than conventional fine-tuning. Our results suggest that learned logit bias is a lightweight mechanism for adapting language models under minimal access requirements.

cs.LG↗

arXiv2026

BayesBench: Evaluating LLM Belief Trajectories Under Multi-Turn Evidence Accumulation

Large language models (LLMs) are typically deployed in multi-turn conversations, where each turn provides new evidence that should reduce epistemic uncertainty about their environment. Acting rationally then requires inferring the unobserved quantities that govern it and updating beliefs about them as evidence accumulates. Yet most evaluations only score the model's final-turn answer in a single-turn format, leaving this process unexamined. We ask how closely LLMs' belief updates match those of a rational Bayesian reasoner in multi-turn settings, and introduce BayesBench, a suite of simulation environments that probe this across three progressively complex tasks: (i) Bayesian estimation, where the model infers an unknown parameter from sequential evidence; (ii) Bayesian prediction, where the model turns inferred beliefs about a latent variable into outcome forecasts; and (iii) latent-framed Bayesian prediction, where observations are filtered through a user-persona framing, requiring joint inference over the latent state and the persona. Across seven LLMs (3B--70B), scaling improves latent inference and evidence accumulation, with updates occasionally matching the Bayesian posterior. However, these gains do not reliably carry over to downstream prediction, exposing a gap between inferring latent structure and using it to rationally update beliefs about the target outcome.

cs.AI↗

arXiv2026

Neural Parameter Calibration for Finite-State Mean Field Games

Mean field games efficiently approximate a very large population of strategic agents. While these games can aid the understanding of complex systems, their deployment in real-world settings is challenged by the specification of their parameters: mean field games (MFGs) often involve hidden preferences, constraints, and interactions that can rarely be theoretically derived or directly observed. To address this gap, we present a neural network-based framework for learning parametric, finite-state MFGs from observed population dynamics. To do so, we formulate the parameter calibration as an inverse problem and use implicit differentiation to backpropagate through the games' equilibrium. The resulting approach is fully differentiable and enables us to estimate flexible trajectory-wise parameter paths, including state- and time-dependent specifications without requiring observations of the individual agents' actions or rewards. We provide a proof for the exactness of the gradient computation in a discrete-time formulation. We validate our framework through numerical experiments across four systems of increasing complexity, ranging from synthetic linear-quadratic benchmarks to real-world urban mobility datasets.

cs.GT↗

arXiv2026

Credit Assignment with Resets in Language Model Reasoning

Contemporary reinforcement learning with verifiable reward methods post-train language models on multi-step reasoning by assigning a single outcome reward uniformly across all tokens in a trajectory. Such uniform assignment ignores which steps contributed to success or failure. Improving credit assignment can address this limitation by enabling targeted refinement of faulty reasoning steps, rather than updating entire trajectories uniformly. Resets are one such simple mechanism, enabling more precise credit assignment by returning to an intermediate state and resampling counterfactual continuations, so that outcome differences can be attributed to decisions made at that point. We propose two such methods: Random-Reset Policy Optimization (RRPO), where reset states are drawn randomly from reasoning steps, and Self-Reset Policy Optimization (SRPO), where the model self-localizes the erroneous step in an incorrect trajectory and resets there. We analyze these methods within the Conservative Policy Iteration (CPI) framework. Extending CPI with a credit-assignment oracle that targets improvable states yields provable improvements over random resets. Across models and reasoning benchmarks, SRPO consistently outperforms standard GRPO and RRPO by sampling multiple suffix continuations at a self-localized reset and learning from their rewards, using only the model itself with no external supervision.

cs.AI↗

arXiv2026

Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale

Aligning autonomous agents with human intent remains a central challenge in modern AI. A key manifestation of this challenge is reward hacking, whereby agents appear successful under the evaluation signal while violating the intended objective. Reward hacking has been observed across a wide range of settings, yet methods for reliably measuring it at scale remain lacking. In this work, we introduce a new evaluation paradigm for measuring reward hacking. Whereas prior studies have primarily analyzed it post hoc by inspecting agent trajectories, we instead embed detectable reward hacking opportunities directly into environments. This makes their exploitation verifiable by design, enabling deterministic and automated measurement of whether and how agents exploit such vulnerabilities. We instantiate this approach in $\textit{TextArena}$ and release $\textit{Hack-Verifiable TextArena}$, a testbed in which reward hacking can be measured reliably. Using this benchmark, we analyze reward hacking behavior across language models in diverse environments and settings. We open source the code at https://github.com/MajoRoth/hack-verifiable-environments/.

cs.LG↗

arXiv2026

Structure Enables Effective Self-Localization of Errors in LLMs

Self-correction in language models remains elusive. In this work, we explore whether language models can explicitly localize errors in incorrect reasoning, as a path toward building AI systems that can effectively correct themselves. We introduce a prompting method that structures reasoning as discrete, semantically coherent thought steps, and show that models can localize errors more reliably within this structure than in conventional, unstructured chain-of-thought reasoning. Motivated by how the human brain monitors errors at discrete decision points and resamples alternatives, we introduce Iterative Correction Sampling of Thoughts (Thought-ICS), a self-correction framework. Thought-ICS iteratively prompts the model to generate reasoning one discrete and complete thought at a time--where each thought represents a deliberate decision by the model--creating natural boundaries for precise error localization. Upon verification, the model localizes the first erroneous step, and the system backtracks to generate alternative reasoning from the last correct point. When asked to correct reasoning verified as incorrect by an oracle, Thought-ICS achieves 20-40% self-correction lift. In a completely autonomous setting without external verification, it outperforms contemporary self-correction baselines.

cs.AI↗

arXiv2025

Imbalanced Gradients in RL Post-Training of Multi-Task LLMs

Multi-task post-training of large language models (LLMs) is typically performed by mixing datasets from different tasks and optimizing them jointly. This approach implicitly assumes that all tasks contribute gradients of similar magnitudes; when this assumption fails, optimization becomes biased toward large-gradient tasks. In this paper, however, we show that this assumption fails in RL post-training: certain tasks produce significantly larger gradients, thus biasing updates toward those tasks. Such gradient imbalance would be justified only if larger gradients implied larger learning gains on the tasks (i.e., larger performance improvements) -- but we find this is not true. Large-gradient tasks can achieve similar or even much lower learning gains than small-gradient ones. Further analyses reveal that these gradient imbalances cannot be explained by typical training statistics such as training rewards or advantages, suggesting that they arise from the inherent differences between tasks. This cautions against naive dataset mixing and calls for future work on principled gradient-level corrections for LLMs.

cs.LG↗

arXiv2025

Self-Improvement of Language Models by Post-Training on Multi-Agent Debate

Self-improvement, where models improve beyond their current performance without external supervision, remains a challenge. The core difficulty is sourcing a training signal stronger than what the model itself can currently produce. Majority voting has been shown to provide such a signal by aggregating over multiple samples, helping mitigate some of the inconsistencies in LM reasoning. In this work, we show that multi-agent debate--where models collaborate and exchange reasoning over multiple rounds--provides an even richer signal than single-round majority voting. We introduce Multi-Agent Consensus Alignment (MACA), which uses reinforcement learning (RL) to post-train models to effectively utilize multi-agent debate. We find that preference learning over full reasoning traces, learning to differentiate between majority and minority reasoning, is more effective than binary consensus rewards or SFT-based approaches for leveraging these debate signals. This produces three key improvements: models are (1) better at utilizing the multi-agent debate setting (+26.87% on MATH), (2) individually more accurate (+21.51% on MathQA), and (3) more self-consistent (+27.6% on GSM8K). We also see strong generalization to unseen benchmarks (+16.3% on GPQA, +11.6% on CommonsenseQA).

cs.AI↗

arXiv2025

ICR-RL: Deep Reinforcement Learning via In-Context Regression

Recent advancements in machine learning have largely been driven by foundation models (FMs) trained on large, diverse datasets, enabling them to generalize effectively to new, related tasks. However, extending this paradigm to reinforcement learning (RL), where an agent interacts with an environment to select actions, remains a significant challenge. Most existing approaches train FMs directly on sets of control tasks, but developing diverse RL environments and scaling training across them can be costly and complex. In this study, we explore a simpler alternative approach based on a classical reduction from RL to regression. We demonstrate that a foundation model pre-trained for regression tasks, when used as an in-context regression (ICR) model, can be directly applied to RL problems. Building on this insight, we introduce a gradient-free method, ICR-RL, that requires no additional training and leverages an ICR foundation model to tackle RL tasks. We evaluate our approach by applying the ICR model with the recently proposed TabPFN, which is trained on a wide range of regression tasks. Experiments conducted on the Gymnasium classic-control benchmark indicate that ICR-RL can compete with commonly used methods, including DQN, PPO and TRPO. These results show that ICR foundation models can effectively solve RL tasks without fine-tuning, demonstrating their potential as a foundation for RL-oriented models.

cs.LG↗

arXiv2025

Simple Optimizers for Convex Aligned Multi-Objective Optimization

It is widely recognized in modern machine learning practice that access to a diverse set of tasks can enhance performance across those tasks. This observation suggests that, unlike in general multi-objective optimization, the objectives in many real-world settings may not be inherently conflicting. To address this, prior work introduced the Aligned Multi-Objective Optimization (AMOO) framework and proposed gradient-based algorithms with provable convergence guarantees. However, existing analysis relies on strong assumptions, particularly strong convexity, which implies the existence of a unique optimal solution. In this work, we relax this assumption and study gradient-descent algorithms for convex AMOO under standard smoothness or Lipschitz continuity conditions-assumptions more consistent with those used in deep learning practice. This generalization requires new analytical tools and metrics to characterize convergence in the convex AMOO setting. We develop such tools, propose scalable algorithms for convex AMOO, and establish their convergence guarantees. Additionally, we prove a novel lower bound that demonstrates the suboptimality of naive equal-weight approaches compared to our methods.

cs.LG↗

arXiv2025

Time After Time: Deep-Q Effect Estimation for Interventions on When and What to do

Problems in fields such as healthcare, robotics, and finance requires reasoning about the value both of what decision or action to take and when to take it. The prevailing hope is that artificial intelligence will support such decisions by estimating the causal effect of policies such as how to treat patients or how to allocate resources over time. However, existing methods for estimating the effect of a policy struggle with \emph{irregular time}. They either discretize time, or disregard the effect of timing policies. We present a new deep-Q algorithm that estimates the effect of both when and what to do called Earliest Disagreement Q-Evaluation (EDQ). EDQ makes use of recursion for the Q-function that is compatible with flexible sequence models, such as transformers. EDQ provides accurate estimates under standard assumptions. We validate the approach through experiments on survival time and tumor growth tasks.

cs.LG↗

arXiv2025

Aligned Multi Objective Optimization

To date, the multi-objective optimization literature has mainly focused on conflicting objectives, studying the Pareto front, or requiring users to balance tradeoffs. Yet, in machine learning practice, there are many scenarios where such conflict does not take place. Recent findings from multi-task learning, reinforcement learning, and LLMs training show that diverse related tasks can enhance performance across objectives simultaneously. Despite this evidence, such phenomenon has not been examined from an optimization perspective. This leads to a lack of generic gradient-based methods that can scale to scenarios with a large number of related objectives. To address this gap, we introduce the Aligned Multi-Objective Optimization framework, propose new algorithms for this setting, and provide theoretical guarantees of their superior performance compared to naive approaches.

cs.LG↗

arXiv2024

Exploiting Structure in Offline Multi-Agent RL: The Benefits of Low Interaction Rank

We study the problem of learning an approximate equilibrium in the offline multi-agent reinforcement learning (MARL) setting. We introduce a structural assumption -- the interaction rank -- and establish that functions with low interaction rank are significantly more robust to distribution shift compared to general ones. Leveraging this observation, we demonstrate that utilizing function classes with low interaction rank, when combined with regularization and no-regret learning, admits decentralized, computationally and statistically efficient learning in offline MARL. Our theoretical results are complemented by experiments that showcase the potential of critic architectures with low interaction rank in offline MARL, contrasting with commonly used single-agent value decomposition architectures.

cs.LG↗

arXiv2024

RL in Latent MDPs is Tractable: Online Guarantees via Off-Policy Evaluation

In many real-world decision problems there is partially observed, hidden or latent information that remains fixed throughout an interaction. Such decision problems can be modeled as Latent Markov Decision Processes (LMDPs), where a latent variable is selected at the beginning of an interaction and is not disclosed to the agent. In the last decade, there has been significant progress in solving LMDPs under different structural assumptions. However, for general LMDPs, there is no known learning algorithm that provably matches the existing lower bound (Kwon et al., 2021). We introduce the first sample-efficient algorithm for LMDPs without any additional structural assumptions. Our result builds off a new perspective on the role of off-policy evaluation guarantees and coverage coefficients in LMDPs, a perspective, that has been overlooked in the context of exploration in partially observed environments. Specifically, we establish a novel off-policy evaluation lemma and introduce a new coverage coefficient for LMDPs. Then, we show how these can be used to derive near-optimal guarantees of an optimistic exploration algorithm. These results, we believe, can be valuable for a wide range of interactive learning problems beyond LMDPs, and especially, for partially observed environments.

cs.LG↗

arXiv2024

Generalizing Multi-Step Inverse Models for Representation Learning to Finite-Memory POMDPs

Discovering an informative, or agent-centric, state representation that encodes only the relevant information while discarding the irrelevant is a key challenge towards scaling reinforcement learning algorithms and efficiently applying them to downstream tasks. Prior works studied this problem in high-dimensional Markovian environments, when the current observation may be a complex object but is sufficient to decode the informative state. In this work, we consider the problem of discovering the agent-centric state in the more challenging high-dimensional non-Markovian setting, when the state can be decoded from a sequence of past observations. We establish that generalized inverse models can be adapted for learning agent-centric state representation for this task. Our results include asymptotic theory in the deterministic dynamics setting as well as counter-examples for alternative intuitive algorithms. We complement these findings with a thorough empirical study on the agent-centric state discovery abilities of the different alternatives we put forward. Particularly notable is our analysis of past actions, where we show that these can be a double-edged sword: making the algorithms more successful when used correctly and causing dramatic failure when used incorrectly.

cs.LG↗

arXiv2024

The Bias of Harmful Label Associations in Vision-Language Models

Despite the remarkable performance of foundation vision-language models, the shared representation space for text and vision can also encode harmful label associations detrimental to fairness. While prior work has uncovered bias in vision-language models' (VLMs) classification performance across geography, work has been limited along the important axis of harmful label associations due to a lack of rich, labeled data. In this work, we investigate harmful label associations in the recently released Casual Conversations datasets containing more than 70,000 videos. We study bias in the frequency of harmful label associations across self-provided labels for age, gender, apparent skin tone, and physical adornments across several leading VLMs. We find that VLMs are $4-7$x more likely to harmfully classify individuals with darker skin tones. We also find scaling transformer encoder model size leads to higher confidence in harmful predictions. Finally, we find improvements on standard vision tasks across VLMs does not address disparities in harmful label associations.

cs.CV↗