论文检索

Reinforcement Learning · Deep RL

Yanzhe Hu, Yijie Jin, Pengfei Liu, Kai Yu, Zhijie Deng

Diffusion Large Language Models (dLLMs) enable parallel token generation, and their block-wise variants have attracted significant attention. However, existing dLLMs usually exhibit an accuracy–parallelism trade-off, where raising tokens per forward (TPF) via aggressive parallel decoding often degrades task accuracy. To address this, we suggest developing a post-training approach to directly optimize the speed–quality frontier of pre-trained dLLMs. Conceptually, we do not require the model to decode aggressively along all sampling trajectories, but rather to find several highly parallelizable ones that can yield correct results. To this end, we resort to a reinforcement learning paradigm, i.e., LightningRL, to optimize rewards regarding both the final accuracy and inference parallelism. LightningRL follows the Group Relative Policy Optimization (GRPO) framework, with further improvements for dLLMs: 1) stabilized training via per-reward decoupled normalization, 2) token-level negative log-likelihood (NLL) loss on correct trajectories for regularization, and 3) improved training efficiency through dynamic sampling with TPF-aware filtering. Across maths and code tasks, LightningRL consistently advances the Pareto frontier, maintaining competitive accuracy while increasing parallelism to an average TPF of 7.3 (up to 11.10 on MBPP).

Reinforcement Learning · Deep RL

Quang Anh PHAM, Tien Mai, Akshat Kumar

Imitation Learning (IL) has demonstrated strong capabilities in learning high-quality policies from expert demonstrations for sequential decision-making tasks. Nonetheless, its effectiveness is significantly constrained in low-expert-data regimes. To mitigate this issue, previous works introduce ``*offline IL with supplementary data*" which augments expert demonstrations with additional, low-cost data generated by suboptimal policies. A prominent framework for this setting is Distribution Correction Estimation (DICE), which estimates the optimal density ratio by solving the dual of a divergence minimization problem between the learned policy and the expert visitation distribution. Despite their theoretical appeal, existing DICE-based methods often require adding a dataset regularizer to the divergence objective, or rely on a strict coverage assumption. These weaknesses limit the capability of DICE-based methods, causing them to be inefficient in some contexts. In this paper, we introduce ReDICE, a new method to address these limitations. ReDICE is derived by formulating an objective under a mixture distribution from the KL divergence between expert and learned policy occupancies. We formally prove that the dual of this formulation is mathematically equivalent to a stable Gumbel regression objective. Furthermore, we introduce a novel policy extraction mechanism that significantly improves performance in practice. Experiments across diverse benchmarks show that ReDICE achieves state-of-the-art results relative to prior offline IL baselines.

Reinforcement Learning · Deep RL

Yunzhe Qi, Sirui Chen, Jiaru Zou, Jingrui He

Effective demonstration selection is crucial for maximizing large language model (LLM) performance in few-shot in-context learning. Due to influences such as recency bias, the effectiveness of demonstrations depends heavily on their context relationship to the specific query, and on the ordering in which they are presented, making demonstration selection a complex combinatorial problem. To address these two challenges, we introduce AutoSelect, a novel framework that formulates demonstration selection as an auto-regressive sequential decision process. At each step, AutoSelect embeds the query and previously selected demonstrations into matrix representations to preserve structural information, and a trainable policy model sequentially selects the next best exemplar. To navigate the factorial space of demonstration permutations, our framework formulates a Kullback-Leibler (KL) regularized optimization problem, from which an optimal policy induces an optimal Plackett-Luce (PL) ranking over all possible demonstration sequences. We prove that minimizing a tractable policy-level Cross-Entropy (CE) loss provably bounds the worst-case discrepancy between our policy's induced PL ranking and the optimal one, enabling tractable prioritization of high-quality sequences. Empirically, AutoSelect outperforms existing heuristic and learning-based methods across nine diverse datasets, achieving up to an 11\% improvement over the strongest baseline. Our results are further supported by analytical studies and a case study, highlighting AutoSelect's key properties, as well as its transferability and generalizability.

Reinforcement Learning · Deep RL

Kazuki Ota, Takayuki Osa, Motoki Omura, Tatsuya Harada

Two-player games such as board games have long been used as traditional benchmark for reinforcement learning. This work revisits a regularized policy optimization with reverse Kullback-Leibler divergence and entropy divergence and analyzes this combination on two-player zero-sum settings from theoretical and empirical perspectives. From a theoretical perspective, we investigate the stability of the policy update rule on two theoretical settings: game-theoretic normal-form games and finite-length games. We provide convergence guarantees and verify our theoretical results by numerical experiments on synthetic games. From an empirical perspective, we derive a practical model-free reinforcement learning algorithm based on the regularized policy optimization. We validate the efficiency in training of our algorithm through comprehensive experiments on five board games: Animal Shogi, Gardner Chess, Go, Hex, and Othello. The experimental results demonstrate that our agent achieves more efficient learning than existing methods across the environments.

Deep Learning · Large Language Models

Alexey Gorbatovski, Boris Shaposhnikov, Viacheslav Sinii, Alexey Malakhov, Daniil Gavrilov

Direct Alignment Algorithms (DAAs) simplify LLM alignment by directly optimizing policies, bypassing reward modeling and RL. While DAAs differ in their use of SFT (one-stage vs. two-stage) and the scalar score they optimize (likelihood vs. odds ratios), the key performance drivers remain underexplored. We present a systematic comparison and analyze a previously overlooked axis - the ranking objective (pairwise vs. pointwise). To isolate this factor, we propose a unified training framework across DAAs by (i) converting one-stage methods (ORPO, ASFT) into a two-stage pipeline with an explicit SFT phase and (ii) introducing a $\beta$ parameter that places all methods in the same hyperparameter space and improves the quality of odds-ratio DAAs (ORPO, ASFT). Under this setup, the ranking objective emerges as the primary determinant of alignment quality, whereas the particular scalar score (policy–reference ratio vs. odds ratio) is secondary. We corroborate this on instruction-following tasks and further confirm it on math-reasoning benchmarks across model scales. Evidence suggests that this stems from how these objectives interact with prompt-specific biases, supported both by strictly controlled experiments and by observations on real data. Our findings underscore the need for nuanced evaluations in DAA research to avoid oversimplified claims of superiority.

Deep Learning · Large Language Models

Tong Wu, Michael Liu, Jun Bai, Zixia Jia, Shuyi Zhang, Ziyong Lin, Yanting Wang, Song-Chun Zhu, Zilong Zheng

We introduce **Native Parallel Reasoner (NPR)**, a teacher-free framework that enables Large Language Models (LLMs) to self-evolve genuine parallel reasoning capabilities. NPR transforms the model from sequential emulation to native parallel cognition through three key innovations: 1) a **self-distilled** progressive training paradigm that transitions from ``cold-start'' format discovery to strict topological constraints without external supervision; 2) a novel **Parallel-Aware Policy Optimization (PAPO)** algorithm that optimizes branching policies directly within the execution graph, allowing the model to learn adaptive decomposition via trial and error; and 3) a robust **NPR Engine** that refactors memory management and flow control of SGLang to enable stable, large-scale parallel RL training. Across eight reasoning benchmarks, NPR trained on Qwen3-4B achieves performance gains of up to 24.5\% and inference speedups up to 4.6$\times$. Unlike prior baselines that often fall back to autoregressive decoding, NPR demonstrates 100\% genuine parallel execution, establishing a new standard for self-evolving, efficient, and scalable agentic reasoning.

Deep Learning · Large Language Models

(Andrew) Zhanke Zhou, Xiangyu Lu, Chentao Cao, Brando Miranda, Tongliang Liu, Bo Han, Sanmi Koyejo

RL with verifiable rewards can substantially improve LLM reasoning, yet standard GRPO-style training often uses uniform sampling and near-uniform weighting, leading to inefficient computation allocation. We study GRPO by tracking token log-probabilities, group-normalized advantages, and induced token-level update weights. This reveals three recurring dynamics: probability inflation, advantage contraction as accuracy rises, and hierarchical convergence, where easy questions quickly saturate while hard questions remain discovery-limited due to rare correct rollouts. These findings imply that the benefit of each update depends strongly on both question difficulty and the model’s current competence. Motivated by this, we propose Confidence and Difficulty-adaptive Policy Optimization (CoDaPO), which assigns each question a bounded value from rollout confidence and empirical difficulty, then uses it to reweight policy updates and resample high-value questions within minibatches to increase discovery under a fixed compute budget. Across seven benchmarks, CoDaPO consistently improves accuracy over other RL methods.

Reinforcement Learning · Deep RL

Po-Nien Kung, Zhen Yang, Jeffrey Luo, Cheng-Fu Yang, Haikang Deng, Zi-Yi Dou, Yinfei Yang, Nanyun Peng, Zhe Gan, Kai-Wei Chang

Large language models can exhibit emergent reasoning behaviors, often manifested as recurring lexical patterns (e.g., “wait,” indicating verification). However, complex reasoning trajectories remain sparse in unconstrained sampling, and standard RL often fails to guarantee the acquisition of diverse reasoning behaviors. We propose a systematic discovery and reinforcement of diverse reasoning patterns through structured reasoning, a paradigm that requires targeted exploration of specific reasoning patterns during the RL process. To this end, we propose Ctrl-R, a framework for learning structured reasoning via tractable trajectory control that actively guides the rollout process, incentivizing the exploration of diverse reasoning patterns that are critical for complex problem-solving. The resulting behavior policy enables accurate importance-sampling estimation, supporting unbiased on-policy optimization. We further introduce a power-scaling factor on the importance-sampling weights, allowing the policy to selectively learn from exploratory, out-of-distribution trajectories while maintaining stable optimization. Experiments demonstrate that Ctrl-R enables effective exploration and internalization of previously unattainable reasoning patterns, yielding consistent improvements across language and vision–language models on mathematical reasoning tasks.

Reinforcement Learning · Deep RL

Daniel Ebi, Gaspard Lambrechts, Damien Ernst, Klemens Böhm

Asymmetric actor-critic methods are widely used in partially observable reinforcement learning, but typically assume full state observability to condition the critic during training, which is often unrealistic in practice. We introduce the informed asymmetric actor-critic framework, allowing the critic to be conditioned on arbitrary state-dependent privileged signals without requiring access to the full state. We show that any such privileged signal yields unbiased policy gradient estimates, substantially expanding the set of admissible privileged information. This raises the problem of selecting the most adequate privileged information in order to improve learning. For this purpose, we propose two novel informativeness criteria: a dependence-based test that can be applied prior to training, and a criterion based on improvements in value prediction accuracy that can be applied post-hoc. Empirical results on partially observable benchmark tasks and synthetic environments demonstrate that carefully selected privileged signals can match or outperform full-state asymmetric baselines while relying on strictly less state information.

Deep Learning · Large Language Models

Lingjing Kong, Xin Liu, Guangyi Chen, Martin Q. Ma, Xiangchen Song, Yuekai Sun, Mikhail Yurochkin, Taylor W. Killian, Russ Salakhutdinov, Kun Zhang 等

Reinforcement learning (RL) has emerged as a key mechanism for transforming LLMs into robust reasoners. While supervised fine-tuning (SFT) often limits models to the distribution of observed reasoning traces, RL post-training significantly improves performance on out-of-distribution (OOD) tasks that require unfamiliar recombinations of familiar steps. We argue that this improvement is driven by **compositional generalization**, which we formalize through a **Hierarchical Latent Selection Model**. In this framework, reasoning traces are generated by a cascade of discrete latent selection variables corresponding to reusable atomic modules, including both skills (local operations) and routing mechanisms (how intermediate information is selected, reused, and composed). We theoretically show that RL’s exploratory nature provides sufficient coverage to identify latent structure and enable compositional generalization. We design controlled experiments to validate this theory. Our results demonstrate that RL can extract atomic modules from compound traces and recombine them to solve new configurations. Moreover, we find that training on compound traces can yield stronger generalization than training on isolated atomic modules. Finally, we investigate relations between SFT and RL and identify an effective protocol in which SFT ensures coverage of all atomic modules, while RL focuses on novel compositions beyond the SFT support to encourage exploration.

Deep Learning · Large Language Models

Xueru Wen, Jie Lou, Yanjiang Liu, Hongyu Lin, Ben He, Xianpei Han, Le Sun, Yaojie Lu, Debing Zhang

While reinforcement learning has achieved impressive progress in language model reasoning, it is constrained by the requirement for verifiable rewards. Recent verifier-free RL methods address this limitation by utilizing the probabilities that LLMs generate reference answers as reward signals. However, these approaches typically sample reasoning traces conditioned only on the question. This design decouples reasoning-trace sampling from answer information, leading to inefficient exploration and incoherence between traces and final answers. In this paper, we propose \textit{\b{Co}upled \b{V}ariational \b{R}einforcement \b{L}earning} (CoVRL), which bridges variational inference and reinforcement learning by coupling prior and posterior distributions through a hybrid sampling strategy. By constructing and optimizing a composite distribution that integrates these two distributions, CoVRL enables efficient exploration while preserving strong thought-answer coherence. Extensive experiments on mathematical and general reasoning benchmarks show that CoVRL improves performance by 12.4\% over the base model and achieves an additional 2.3\% improvement over state-of-the-art verifier-free RL baselines, providing a principled framework for enhancing the general reasoning capabilities of language models.

Reinforcement Learning · Deep RL

Yiming Wang, Yiheng Zhang, Kaiyan Zhao, Xingjie Zuo, Xingyu Liu, Xuetao Li, Furui Liu, Bo An, Leong Hou U

Efficient exploration remains a key challenge in deep reinforcement learning, especially for embodied agents operating in realistic environments with high-dimensional observations and complex dynamics. Recent latent exploration methods define bonuses in a learned latent space, but often struggle in these settings where (i) representations can be noisy or policy-dependent, and (ii) common strategies such as randomized latent objectives or fixed directional spanning are brittle and fail to improve global coverage. We propose Covariance Volume Maximization (CVM), a coverage-driven latent exploration framework with two key components. First, we learn a behavioral state encoder using a *policy-mixture* objective to reduce representation drift under rapidly changing exploration policies, yielding stable and behaviorally meaningful latent displacements. Second, CVM rewards each transition by its exact increase in the log-determinant of the covariance of recent latent displacements, explicitly expanding the explored region and prioritizing under-covered directions. This objective coincides with the classical *D-optimal* design criterion, providing an information-efficiency justification. Extensive experiments on embodied navigation and manipulation tasks demonstrate that CVM substantially improves exploration efficiency and robustness, and scales effectively to different environments.

Reinforcement Learning · Deep RL

Pengcheng Jiang, Judith Li, Moonkyung Ryu, Lily Hu, Kun Su, Zhong Yi Wan, Liam Hebert, Hao Peng, Jiawei Han, Dima Kuzmin 等

Many modern retrieval problems are \emph{set-valued}: given a broad intent, the system must return a \emph{collection} of results that optimizes higher-order properties (e.g., diversity, coverage, complementarity, coherence) while staying grounded to a fixed database. Set-valued objectives are inherently non-decomposable and are not captured by existing supervised (query, content) datasets which only prioritize top-1 retrieval. While reinforcement learning (RL) can optimize set-level objectives via interaction, deploying an RL-tuned LLM for fan-out retrieval is prohibitively expensive at query time. Conversely, diffusion-based generative retrieval enables efficient single-pass fan-out in embedding space, but requires objective-aligned training targets. To address these issues, we propose {R4T (Retrieve-for-Train)}, which uses RL \emph{once} as an objective transducer in a three step process: (i) train a fan-out LLM with composite set-level rewards, (ii) synthesize objective-consistent training pairs, and (iii) train a lightweight diffusion retriever to model the conditional distribution of set-valued outputs. Across Polyvore and a music playlist dataset, R4T improves retrieval quality over strong baselines while reducing query-time fan-out latency by an order of magnitude.

Reinforcement Learning · Deep RL

John Wikman, Alexandre Proutiere, David Broman

In standard reinforcement learning (RL) settings, the interaction between the agent and the environment is typically modeled as a Markov decision process (MDP), which assumes that the agent observes the system state instantaneously, selects an action without delay, and executes it immediately. In real-world dynamic environments, such as cyber-physical systems, this assumption often breaks down due to delays in the interaction between the agent and the system. These delays can vary stochastically over time and are typically _unobservable_ when deciding on an action. Existing methods deal with this uncertainty conservatively by assuming a known fixed upper bound on the delay, even if the delay is often much lower. In this work, we introduce the _interaction layer_, a general framework that enables agents to adaptively handle unobservable and time-varying delays. Specifically, the agent generates a matrix of possible future actions, anticipating a horizon of potential delays, to handle both unpredictable delays and lost action packets sent over networks. Building on this framework, we develop a model-based algorithm, _Actor-Critic with Delay Adaptation (ACDA)_, which dynamically adjusts to delay patterns. Our method significantly outperforms state-of-the-art approaches across a wide range of locomotion benchmark environments, including real-world measured delays.

Reinforcement Learning · Deep RL

Tsunehiko Tanaka, Kenshi Abe, Kaito Ariu, Tetsuro Morimura, Edgar Simo-Serra

Traditional approaches in offline reinforcement learning aim to learn the optimal policy that maximizes the cumulative reward, also known as return. It is increasingly important to adjust the performance of AI agents to meet human requirements, for example, in applications like video games and education tools. Decision Transformer (DT) optimizes a policy that generates actions conditioned on the target return through supervised learning and includes a mechanism to control the agent's performance using the target return. However, the action generation is hardly influenced by the target return because DT’s self-attention allocates scarce attention scores to the return tokens. In this paper, we propose Return-Aligned Decision Transformer (RADT), designed to more effectively align the actual return with the target return. RADT leverages features extracted by paying attention solely to the return, enabling action generation to consistently depend on the target return. Extensive experiments show that RADT significantly reduces the discrepancies between the actual return and the target return compared to DT-based methods.

Reinforcement Learning · Deep RL

Xiwen Chen, Wenhui Zhu, Jingjing Wang, Peijie Qiu, Zhipeng Wang, Huayu Li, ZhengXiao He, XUANZHAO DONG, Prayag Tiwari, Mingkun Xu 等

Aligning Large Language Models (LLMs) with human preferences is often formulated via Direct Preference Optimization (DPO). However, the standard Bradley-Terry instantiation of DPO is limited in modeling common departures from transitivity in human preferences. To address this, recent work has introduced Self-Play Preference Optimization (SPPO), which iteratively refines the policy by training on self-generated win-lose pairs. Our investigation, however, reveals a critical instability in SPPO: the optimization is prone to \textit{policy degeneration} when the preference oracle assigns overly confident wins to semantically indistinguishable responses. To mitigate this, we propose $\textit{S}$-SPPO, a dual-space semantic calibration framework comprising: i) $\textit{Supervision Calibration}$ via semantic gating, which anneals win rate targets toward the maximum-entropy baseline as semantic overlap increases; and ii) $\textit{Representation Calibration}$ via latent repulsion to enforce geometric diversity to prevent manifold collapse and maintain latent diversity between chosen and rejected samples. Theoretically, we show that the calibration preserves the constant-sum game structure, facilitating convergence to a Nash Equilibrium. Empirically, $\textit{S}$-SPPO avoids the performance degradation seen in prior methods, achieving 52.19\% win rate and 47.46\% length-controlled win rate on AlpacaEval 2.0 with Llama-3-8B, without using additional human-annotated preferences during training.

Deep Learning · Large Language Models

Ian Wu, Yuxiao Qu, Amrith Setlur, Aviral Kumar

Large Language Models (LLMs) that continue improving at test-time budgets far beyond their training budgets can solve harder problems by leveraging additional inference compute: we refer to this property as extrapolation. Standard on-policy RL operates on fixed problem distributions and training budgets, giving rise to a distribution shift between train and test that limits the resulting model's extrapolation capabilities. To address this, we introduce RC, an iterative decoding algorithm replacing standard autoregressive decoding that enables models to extrapolate to lengths an order of magnitude longer than those seen during training. RC exploits the asymmetry between summarization and generation capabilities present in LLMs to construct a decoding process that improves consistently over iterations. Its effectiveness can be further increased through training, which amplifies the model’s ability to perform summary-conditioned reasoning while avoiding the challenges of long-horizon RL. Empirically, training a 4B instruction-following model with RC using a 16k-token training budget improves performance on HMMT 2025 from 40% to 70% when evaluated with a 512k-token test budget, substantially surpassing comparably sized LLMs.

Deep Learning · Generative Models and Autoencoders

Guanghan Wang, Gilad Turok, Yair Schiff, Marianne Arriola, Volodymyr Kuleshov

While diffusion language models (DLMs) have achieved competitive performance in text generation, improving their reasoning ability with reinforcement learning remains an active research area. Here, we introduce d2, a reasoning framework tailored for masked DLMs. Central to our framework is a new policy gradient algorithm that relies on accurate estimates of the sampling trajectory likelihoods. Our likelihood estimator, d2-AnyOrder, achieves exact trajectory likelihood with a single model pass for DLMs that support a sampling algorithm called any-order decoding. Through an empirical study of widely used DLMs, we show that any-order decoding is not universally supported in practice. Consequently, for DLMs that do not naturally support any-order decoding, we propose another estimator, d2-StepMerge, which, unlike d2-AnyOrder, only approximates the trajectory likelihood. d2-StepMerge trades off compute for approximation accuracy in an analytically tractable manner. Empirically, d2 significantly outperforms widely-used RL baselines when applied to popular DLMs, and sets a new state-of-the-art performance for DLMs on logical reasoning tasks (Countdown and Sudoku) and math reasoning benchmarks (GSM8K and MATH500).

Deep Learning · Large Language Models

Rana Khan, Zijie Liu, Zhen Tan, Charles Fleming, Tianlong Chen

Reinforcement Learning (RL) and Supervised Fine-Tuning (SFT) are the two dominant paradigms for enhancing Large Language Model (LLM) performance on downstream tasks. While RL generally preserves broader model capabilities (retention) better than SFT, it comes with significant costs: complex reward engineering, instability, and expensive on-policy sampling. In contrast, SFT is efficient but brittle, often suffering from catastrophic forgetting due to $\textbf{Supervision Mismatch}$: the divergence between the model's evolving policy and static training labels. We address this trade-off with $\textbf{Trajectory-Mixed Supervision (TMS)}$, a reward-free framework that approximates the on-policy benefits of RL by creating a dynamic curriculum from the model's own historical checkpoints. TMS minimizes $\textit{Policy-Label Divergence (PLD)}$, preventing the mode collapse that drives forgetting in standard SFT. Experiments across reasoning (MATH, GSM8K) and instruction-following benchmarks demonstrate that TMS effectively shifts the accuracy-retention Pareto frontier. While RL remains the gold standard for retention, TMS significantly outperforms standard and iterative SFT, bridging the gap to RL without requiring reward models or verifiers. Mechanistic analysis confirms that PLD drift accurately predicts forgetting, and that TMS successfully mitigates this drift.

Deep Learning · Large Language Models

Isha Puri, Mehul Damani, Idan Shenfeld, Marzyeh Ghassemi, Jacob Andreas, Yoon Kim

Large language models (LMs) are typically post-trained via RL to produce a single best answer per query, implicitly optimizing for modal correctness. While effective for benchmark accuracy, this approach is unideal for many applications of interest such as in medical diagnosis, which would benefit from models generating a set of plausible answers (ideally paired with uncertainty estimates).This paper describes a multi-answer reinforcement learning (RL) approach for enabling LMs to do this, where we modify the RL objective to train models to explicitly generate multiple candidate answers in a single forward pass, internalizing aspects of inference-time search into the model’s generative process. We instantiate this approach through Multi-Answer Reinforcement Learning with Verifiable Rewards (Multi-RLVR), which generalizes ordinary RLVR to the multi-answer case with a set-level reward. We further extend this approach to Multi-Answer Reinforcement Learning with Calibrated Rewards (Multi-RLCR) which adds a set-level Brier score-based calibration objective to enable LMs to output calibrated uncertainty estimates associated with each answer in the output set. Multi-answer training promotes explicit representation of alternative hypotheses rather than repeated generation of the dominant mode. Across question-answering and medical diagnostic benchmarks, we observe improved diversity, recall, and set-level calibration scores compared to single answer-trained baselines. We further observe that models trained with our approach are more token-efficient, requiring fewer tokens to generate multiple answers than competing approaches. These results position multi-answer RL as a principled and compute-efficient alternative to inference-time scaling.