论文检索

Deep Learning · Large Language Models

Sangeun Park, Minhae Kwon

A central goal of large language model (LLM) research is to build agentic systems that can plan, act, and adapt through sustained interaction with dynamic environments. While recent LLM-based agents exhibit impressive contextual reasoning, their long-horizon decision-making remains fragile, often suffering from $\textit{objective drift}$, where goals and plans drift over extended interactions. We introduce $\texttt{Multi}^2$, a hierarchical multi-agent decision-making framework that explicitly decomposes agent behavior into complementary roles. A high-level agent ($\texttt{System 1}$) focuses on context-aware sub-goal generation using supervised fine-tuning (SFT), while a low-level agent ($\texttt{System 2}$) executes atomic actions through offline-to-online reinforcement learning (RL) in interactive environments. This separation enables stable long-horizon control, mitigates objective drift, and allows efficient adaptation. Across diverse interactive environments, $\texttt{Multi}^2$ consistently outperforms strong agentic baselines, demonstrating improved robustness and coordination in multi-turn interaction. Beyond performance, we introduce and release three hierarchical benchmark datasets, filling a long-standing gap in training and evaluating hierarchical decision-making for LLM-based agents.

Deep Learning · Large Language Models

Linzheng Chai, Jian Yang, Jiajun Wu, Ensheng Shi, Xianglong Liu

Current reinforcement learning (RL) methods for code generation are predominantly optimized on Python, showing weak generalization to other programming languages (PLs). Although leveraging multilingual solutions offers richer semantics and a wider search landscape, naive independent training across languages suffers from optimization imbalance and fails to effectively transfer knowledge from high-resource languages. We propose Group Cross-lingual Relative Policy Optimization (GXPO), which forms training groups by generating solutions for the same problem in multiple PLs and jointly optimizes language-specific and cross-language signals, enabling more balanced optimization and improved transfer to low-resource PLs. We additionally introduce Multilingual LiveCodeBench (ML-LCB), extending LiveCodeBench to a unified multilingual evaluation setting. On ML-LCB across 8 PLs, GXPO consistently improves performance, with pronounced gains on low-resource PLs, demonstrating scalable multilingual RL for language-consistent code generation.

Reinforcement Learning · Deep RL

Gwangpyo Yoo, Woo Kyung Kim, Honguk Woo

In this work, we extend distributional reinforcement learning (RL) to develop a risk-sensitive multi-objective RL framework, with applications to domains such as finance and robotics. We achieve this by adopting vector-risk measures and approximating them via Knothe-Rosenblatt (KR) quantile regression. This approach directly extends the IQN framework to the multi-objective setting, aligns with the axiomatic definition of vector-risk measures, and guarantees that critics converge under the distributional Bellman operator. To mitigate the artificial ordering imposed by the KR map, we employ a transformer architecture without positional encoding, and introduce MO-TQC for training stability. We demonstrate improved performance on MO-Gymnasium benchmarks and use our framework to study risk-sensitive policies in multi-objective tasks.

Deep Learning · Large Language Models

Yuntian Tang, Bohan Jia, Wenxuan Huang, Lianyue Zhang, Jiao Xie, Wenxi Li, Wei Li, Jie Hu, Xinghao Chen, Rongrong Ji 等

Chain-of-Thought (CoT) reasoning successfully enhances the reasoning capabilities of Large Language Models (LLMs), yet it incurs substantial computational overhead for inference. Existing CoT compression methods often suffer from a critical loss of logical fidelity at high compression ratios, resulting in significant performance degradation. To achieve high-fidelity, fast reasoning, we propose a novel EXTreme-RAtio Chain-of-Thought Compression framework, termed Extra-CoT, which aggressively reduces the token budget while preserving answer accuracy. To generate reliable, high-fidelity supervision, we first train a dedicated semantically-preserved compressor on mathematical CoT data with fine-grained annotations. An LLM is then fine-tuned on these compressed pairs via a mixed-ratio supervised fine-tuning (SFT), teaching it to follow a spectrum of compression budgets and providing a stable initialization for reinforcement learning (RL). We further propose Constrained and Hierarchical Ratio Policy Optimization (CHRPO) to explicitly incentivize question-solving ability under lower budgets by a hierarchical reward. Experiments on three mathematical reasoning benchmarks show the superiority of Extra-CoT. For example, on MATH-500 using Qwen3-1.7B, Extra-CoT achieves over 73\% token reduction with an accuracy improvement of 0.6\%, significantly outperforming state-of-the-art (SOTA) methods. Our source codes are released in the Supplementaries.

Reinforcement Learning · Deep RL

Jeong Woon Lee, Kyoleen Kwak, Daeho Kim, Hyoseok Hwang

Policies learned via continuous actor-critic methods often exhibit erratic, high-frequency oscillations, making them unsuitable for physical deployment. Current approaches attempt to enforce smoothness by directly regularizing the policy's output. We argue that this approach treats the symptom rather than the cause. In this work, we theoretically establish that policy non-smoothness is fundamentally governed by the differential geometry of the critic. By applying implicit differentiation to the actor-critic objective, we prove that the sensitivity of the optimal policy is bounded by the ratio of the Q-function's mixed-partial derivative (noise sensitivity) to its action-space curvature (signal distinctness). To empirically validate this theoretical insight, we introduce PAVE (Policy-Aware Value-field Equalization), a critic-centric regularization framework that treats the critic as a scalar field and stabilizes its induced action-gradient field. PAVE rectifies the learning signal by minimizing the Q-gradient volatility while preserving local curvature. Experimental results demonstrate that PAVE achieves smoothness and robustness comparable to policy-side smoothness regularization methods, while maintaining competitive task performance, without modifying the actor.

Reinforcement Learning · Deep RL

Yudong Luo, Erick Delage

Optimizing Conditional Value-at-risk (CVaR) using policy gradient (a.k.a CVaR-PG) faces significant challenges of sample inefficiency. This inefficiency stems from the fact that it focuses on tail-end performance and overlooks many sampled trajectories. We address this problem by augmenting CVaR with an expected quantile term. Quantile optimization admits a dynamic programming formulation that leverages all sampled data, thus improves sample efficiency. This does not alter the CVaR objective since CVaR corresponds to the expectation of quantile over the tail. Empirical results in domains with verifiable risk-averse behavior show that our algorithm within the Markovian policy class substantially improves upon CVaR-PG and consistently outperforms other existing methods.

Social Aspects · Accountability, Transparency, and Interpretability

Andrea Marzo, Alessio Ragno, Roberto Capobianco

Reinforcement learning (RL) has achieved remarkable success across complex decision-making tasks, especially with the advent of deep neural networks. However, the resulting models are often opaque, making their deployment in safety-critical domains challenging. Explainable AI aims to address this issue, but most specific efforts for deep RL remain limited either to post-hoc explanation methods or to imitation learning and distillation procedures. These latter approaches rely on pre-trained black-box agents and are typically restricted to environments with discrete action spaces, limiting their scalability and interpretability. In this paper, we introduce ProtoSAC, a novel deep RL architecture that integrates a prototype-based actor into the Soft Actor-Critic (SAC) algorithm, enabling intrinsic interpretability in continuous action spaces. Our method learns a set of prototypes that represent interpretable state clusters, each associated with a Gaussian action distribution. Actions are generated as a similarity-weighted mixture over these prototypes, providing transparent decision-making without sacrificing performance. We evaluate ProtoSAC on continuous action-space environments and show that it matches the performance of the original SAC while offering enhanced interpretability.

Reinforcement Learning · Deep RL

M Ganesh Kumar, Adam Lee, Blake Bordelon, Cengiz Pehlevan

The maximal update parameterization ($\mu P$) has been influential in supervised and unsupervised learning conditions, with fixed data distributions, owing to its ability to maintain feature learning across larger parameter scales. This parameterization facilitates more consistent learning dynamics and learned features across model sizes. Moreover, optimal hyperparameters such as learning rate approximately transfer from small to larger models, minimizing the computational overhead of hyperparameter sweeps. However, it remains elusive if these benefits readily transfer to the reinforcement learning framework, where the model's learning dynamics are coupled to the shifting data distribution. Reinforcement learning agents must continually adapt to non-stationary data distribution shifts throughout training. We empirically study how two regimes, the ''rich'' CompleteP and ''lazy'' Neural Tangent Kernel (NTK) parameterizations affect hyperparameter transfer, feature and policy consistency as we scale reinforcement learning agents. Ultimately, we show that agents trained using CompleteP consequentially improves compute and reward efficiency compared to the NTK parameterization over 16 continuous control tasks and variants e.g. normalization and sparse rewards. Hence, we argue that adopting the CompleteP parameterization minimizes learning inconsistencies across model sizes to improve compute efficiency when scaling up.

Deep Learning · Large Language Models

Mahesh Ramesh, Kaousheik Jayakumar, Aswinkumar Ramkumar, Pavan Thodima, Aniket Rege, Emmanouil-Vasileios Vlatakis-Gkaragkounis

Cooperative reasoning under incomplete information remains complex for both humans and multi-agent AI, requiring agents to transcend individual logic in favor of recursive Theory-of-Mind (ToM) and strategic coordination. To investigate these challenges, we conduct a large-scale evaluation of 17 state-of-the-art LLMs (4B–600B+) on Hanabi card game across 2–5 players. To examine their limitations, we analyze the impact of context engineering and scaffold robustness, ranging from minimal prompts (Watson setting) to Bayesian-motivated scaffolding (Sherlock setting) and multi-turn working memory (Mycroft setting). Our findings reveal that: (1) top-performing models can autonomously track game states via internal working memory, although not reliably, and (2) cross-play performance scales smoothly with model capability. However, even the best models (scoring ≈ 15/25) trail specialist human experts (> 20/25). We introduce and release two novel datasets: HanabiLogs (1,520 annotated trajectories) and HanabiRewards (560 games with dense move-level utilities). By fine-tuning a 4B open-weight model (Qwen3-Instruct) on our datasets, we achieve performance gains of up to 156%, bringing performance to within 3 points of a strong proprietary reasoning model (o4-mini) and surpassing the best non-reasoning model (GPT-4.1) by 52%. Crucially, our HanabiRewards RL-finetuned model further generalizes beyond Hanabi, improving performance on a cooperative group-guessing benchmark by 11%, temporal reasoning on EventQA by 6.4%, instruction-following on IFBench-800K by 1.7 Pass@10, and matching AIME 2025 mathematical reasoning Pass@10. Code and datasets are available at {redacted for double blind}.

Deep Learning · Large Language Models

Zizhe Chen, Jiqian Dong, Yizhou Tian, Garry YANG, Yongqiang Chen, Zhitang Chen, James Cheng

Reinforcement Learning (RL) refines large language models (LLMs) by directly optimizing model behavior with reward signals. Although accurate state value estimation is essential for stable training in classical RL settings, it remains an understudied challenge in LLM post-training. In this work, we demonstrate that accurate value estimation can stabilize and improve post-training. First, we construct State Value Estimation Benchmark (SVEB) and show that critics of standard approaches like PPO simply degenerate toward a coarse group-average baseline. To overcome this, we propose two techniques. One is a heuristic method *Numca*, which uses numbers in responses as state representation to calculate state value. Another is a general hidden-state-based framework *Hista*, which utilize the semantic information in hidden states to group disjoint responses. Experiments show that when equipped with these improved estimates, training gains better performance consistently with different RL algorithms.

Deep Learning · Large Language Models

Dulhan Jayalath, Shashwat Goel, Thomas Foster, Parag Jain, Suchin Gururangan, Cheng Zhang, Anirudh Goyal, Alan Schelten

Where do learning signals come from when there is no ground truth in post-training? We show that inference compute itself can serve as supervision. By generating parallel rollouts and converting them into reference estimates, models can learn without human labels—critically, even in non-verifiable domains like healthcare guidance where no programmatic checker exists. We call this framework *Compute as Teacher (CaT)* and it turns inference-time compute from parallel rollouts into supervision for RL training. The framework has two components: (1) reference estimation which aggregates rollouts into a pseudo-reference answer, and (2) reward derivation which converts that pseudo-reference into RL rewards. For (1), we explore a simple method we call *synthesis*, but the framework admits any aggregator. For (2), we introduce self-proposed rubrics for non-verifiable domains. These are binary, auditable criteria generated from the pseudo-reference and scored by an LLM judge. On HealthBench, models trained with CaT match or exceed inference-time aggregation quality while using 9× less test-time compute. Here, CaT also competes with learning from expert physician annotations, yielding up to +30% relative improvement over the initial policy. The framework extends naturally to verifiable rewards, matching the best existing baselines on MATH-500 in test-time RL and demonstrating 'drop-in' versatility across both types of domains.

Deep Learning · Large Language Models

Xiaozhe Li, Yang Li, Xinyu Fang, Shengyuan Ding, Peiji Li, Yongkang Chen, Yichuan Ma, TianYi Lyu, Linyang Li, Dahua Lin 等

On-policy reinforcement learning methods like GRPO suffer from \emph{mode collapse}: they exhibit reduced solution diversity, concentrating probability mass on a single solution once discovered and ceasing exploration of alternative strategies. We show this stems from reverse KL minimization's mode-seeking behavior, which reinforces the first high-reward trajectory found rather than maintaining a distribution over multiple diverse solutions. We propose DMPO (\textbf{D}istribution-\textbf{M}atching \textbf{P}olicy \textbf{O}ptimization), which prevents mode collapse through principled approximation of forward KL minimization. DMPO constructs a group-level target distribution over sampled trajectories proportional to their rewards, then aligns the policy distribution to this target. This provides mode-covering behavior without requiring sampling from the intractable global target distribution, enabling sustained exploration throughout training. We validate DMPO on NP-hard combinatorial optimization, where exponentially many feasible solutions exist but only a few approach optimality—an ideal testbed for evaluating exploration. DMPO achieves {43.9\% Quality Ratio on text-based NP-Bench (vs. GRPO's 40.1\%)} and {43.1\% on vision-based NP-Bench (vs. 38.4\%)}—demonstrating 9\% and 12\% relative improvements respectively. These gains generalize to mathematical reasoning (+2.0\%) and out-of-domain tasks (+2.3\%), showing that diversity-preserving training enhances general reasoning capabilities across modalities. Our work establishes distribution matching as a practical, principled approach to preventing mode collapse in on-policy RL, with consistent quality improvements demonstrating sustained exploration across diverse reasoning tasks.

Deep Learning · Large Language Models

Yuxiang Wei, Zhiqing Sun, Emily McMilin, Jonas Gehring, David Zhang, Gabriel Synnaeve, Daniel Fried, LINGMING ZHANG, Sida Wang

While current software agents powered by large language models (LLMs) and reinforcement learning (RL) can boost programmer productivity, their reliance on human-curated training data and environments creates a fundamental barrier to superintelligence. In this paper, we present Self-play SWE-RL (SSR), a first step toward training superintelligent software agents under minimal data assumptions. SSR requires only access to sandboxed repositories with source code and dependencies, no need for human-labeled is sues or test commands. Grounded in real-world codebases, a single LLM agent is trained via RL in a self-play setting to inject and repair increasingly complex bugs. The bugs are formally specified by test suite improvements proposed by the agent rather than natural language issue descriptions. On the SWE-bench Verified and SWE-Bench Pro benchmarks, SSR achieves clear self-improvement (+10.4 and +7.8 points) and consistently outperforms the human-data baseline throughout training, generalizing to natural language bug descriptions not seen in training. Overall, our results point toward a paradigm where agents autonomously gather extensive learning experiences from real software repositories, ultimately enabling superintelligent systems that exceed human capabilities in understanding, modifying, and creating software from scratch.

Deep Learning · Foundation Models

David Schiff, Ofir Lindenbaum, Yonathan Efroni

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 matches or outperforms state-of-the-art methods, including DQN and PPO. 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

Deep Learning · Large Language Models

Bowei He, Minda Hu, Zenan Xu, Hongru WANG, Licheng Zong, Yankai Chen, Chen Ma, Xue Liu, PlutoZhou, Irwin King

Search-integrated reasoning enables language agents to transcend static parametric knowledge by actively querying external sources. However, training these agents via reinforcement learning is hindered by the *multi-scale credit assignment* problem: existing methods typically rely on sparse, trajectory-level rewards that fail to distinguish between high-quality reasoning and fortuitous guesses, leading to redundant or misleading search behaviors. To address this, we propose Search-R2, a novel Actor–Refiner collaboration framework that enhances reasoning through targeted intervention, with both components jointly optimized during training. Our approach decomposes the generation process into an Actor, which produces initial reasoning trajectories, and a Meta-Refiner, which selectively diagnoses and repairs flawed steps via a ``cut-and-regenerate'' mechanism. To provide fine-grained supervision, we introduce a hybrid reward design that couples outcome correctness with a dense process reward quantifying the information density of retrieved evidence. Theoretically, we formalize the Actor–Refiner interaction as a smoothed mixture policy, proving that selective correction yields strict performance gains over strong baselines. Extensive experiments across various general and multi-hop QA datasets demonstrate that Search-R2 consistently outperforms strong RAG and RL-based baselines across model scales, achieving superior reasoning accuracy with minimal overhead.

Deep Learning · Large Language Models

Shobhita Sundaram, John Quan, Ariel Kwiatkowski, Kartik Ahuja, Yann Ollivier, Julia Kempe

RL methods for finetuning large reasoning models stall on datasets with low initial success rates, and thus little training signal. We investigate a fundamental question: Can a pretrained LLM leverage latent knowledge to generate an automated curriculum for problems it cannot solve? We explore this with SOAR: A self-improvement framework designed to surface these pedagogical signals through meta-RL. A teacher model proposes synthetic problems for a student model, and is rewarded with its improvement on a subset of hard problems, thus grounding the curriculum in real student progress rather than proxy rewards. Our study on the hardest subsets of math benchmarks (0/128 success) reveal three core findings. First, it is possible to realize bi-level meta-RL that unlocks learning under sparse, binary rewards by sharpening a latent capacity of pretrained models to generate useful problems. Second, grounded rewards outperform intrinsic rewards used in prior LLM self-play, reliably avoiding the typical instability and diversity collapse modes. Third, the structure and well-posedness of questions are more critical for learning progress than solution correctness. Our results suggest that the ability to generate useful stepping stones does not require the preexisting ability to solve the hard problems, paving a principled path to escape reasoning plateaus without additional curated data.

Deep Learning · Large Language Models

Charles Arnal, Vivien Cabannnes, Taco Cohen, Julia Kempe, REMI MUNOS

While Experience Replay—the practice of storing rollouts and reusing them multiple times during training—is a foundational technique in general RL, it remains largely unexplored in LLM post-training due to the prevailing belief that fresh, on-policy data is essential for high performance. In this work, we challenge this assumption. We present a systematic study of replay buffers for LLM post-training, formalizing the optimal design as a trade-off between staleness-induced variance, sample diversity and the high computational cost of generation. We show that strict on-policy sampling is suboptimal when generation is expensive. Empirically, we show that a well-designed replay buffer can drastically reduce inference compute without degrading -- and in some cases even improving -- final model performance, while preserving policy entropy.

Reinforcement Learning · Deep RL

Elizabeth Bates, Chris Hicks, Vasilios Mavroudis

Recent years have seen an explosion of interest in autonomous cyber defence agents trained to defend computer networks using deep reinforcement learning. These agents are typically trained in cyber gym environments using dense, highly engineered reward functions which combine many penalties and incentives for a range of (un)desirable states and costly actions. Dense rewards help alleviate the challenge of exploring complex environments but risk biasing agents towards suboptimal and potentially riskier solutions, a critical issue in complex cyber environments. We thoroughly evaluate the impact of reward function structure on learning and policy behavioural characteristics using a variety of sparse and dense reward functions, two well-established cyber gyms, a range of network sizes, and both policy gradient and value-based RL algorithms. Our evaluation is enabled by a novel ground truth evaluation approach which allows directly comparing between different reward functions, illuminating the nuanced inter-relationships between rewards, action space and the risks of suboptimal policies in cyber environments. Our results show that sparse rewards, provided they are goal aligned and can be encountered frequently, uniquely offer both enhanced training reliability and more effective cyber defence agents with lower-risk policies. Surprisingly, sparse rewards can also yield policies that are better aligned with cyber defender goals and make sparing use of costly defensive actions without explicit reward-based numerical penalties.

Deep Learning · Large Language Models

Ma Yan, Weiyu Zhang, Tianle Li, Du Linge, Xuyang Shen, Pengfei Liu

Vision tool-use reinforcement learning (RL) can equip vision--language models with visual operators such as crop-and-zoom and achieves strong performance gains, yet it remains unclear whether these gains are driven by improvements in tool use or evolving intrinsic capabilities. We introduce \textbf{MED} (Measure--Explain--Diagnose), a coarse-to-fine framework that disentangles intrinsic capability changes from tool-induced effects, decomposes the tool-induced performance difference into gain and harm terms, and probes the mechanisms driving their evolution. Across checkpoint-level analyses on two VLMs with different tool priors and six benchmarks, we find that improvements are dominated by intrinsic learning, while tool-use RL mainly reduces tool-induced harm (e.g., fewer call-induced errors and weaker tool schema interference) and yields limited progress in tool-based correction of intrinsic failures.Overall, current vision tool-use RL learns to coexist safely with tools rather than master them.

Reinforcement Learning · Deep RL

Hikmet Simsir, Ozgur S. Oguz

Behavior cloning with high-capacity generative policies achieves strong imitation performance, but performance is often constrained by limited demonstration coverage and sensitivity to distribution shift. While reinforcement learning can improve task performance, directly fine-tuning large action decoders is often unstable and sample inefficient. We propose **Lagrangian Perturbation Diffusion Steering (LP-DS)**, a lightweight adaptation method that improves a frozen generative policy while preserving its multimodal structure. LP-DS learns a compact noise-space perturbation module that shifts Gaussian noise inputs before decoding, enabling policy improvement without modifying the action decoder. To prevent off-manifold latent queries and unstable denoising dynamics, we optimize this module with a Lagrangian trust-region objective that maximizes downstream value while constraining perturbation magnitude, yielding stable and sample-efficient learning. Across RoboMimic manipulation, OpenAI Gym locomotion, and Adroit dexterous manipulation benchmarks, LP-DS improves sample efficiency, success, and return while maintaining diverse behavior, as quantified by higher action-space entropy using the Kozachenko--Leonenko k-nearest neighbor estimator, with return improvements of up to 25\% over prior baselines. Anonymous project page: https://sites.google.com/view/lp-ds/home.