论文检索

Deep Learning · Large Language Models

Jiaqi Wei, Xuehang Guo, Pengfei Yu, Xiang Zhang, Wanli Ouyang, Siqi Sun, Qingyun Wang, Chenyu You

Standard Chain-of-Thought (CoT) reasoning trades reliability for responsiveness: in a single user-visible token stream, more deliberation delays meaningful output, imposing a ``silence tax.'' We introduce \emph{Side-by-Side (SxS) Interleaved Reasoning}, a training framework that makes \emph{disclosure timing} a controllable decision within standard autoregressive generation. By interleaving \emph{supported} partial answers with continued private reasoning in the same context, SxS avoids monolithic reasoning preambles. We treat disclosure as a policy learning problem and train models via a multi-stage pipeline: supervised fine-tuning (SFT) on entailment-aligned interleaved trajectories, followed by reinforcement learning (RL) to recover reasoning performance and optimize accuracy. Without architectural changes, SxS improves the accuracy--latency trade-off across two Qwen3 architectures/scales (MoE \textbf{Qwen3-30B-A3B}, dense \textbf{Qwen3-4B}) and both in-domain (AIME25) and out-of-domain (GPQA-Diamond) benchmarks, reducing \emph{substantive content latency} by 18\% and improving a proxy for perceived wait time by 49\%, yielding more responsive interactions without compromising answer quality.

Reinforcement Learning · Deep RL

Paulius Sasnauskas, Yiğit Yalın, Goran Radanovic

We study the corruption-robustness of in-context reinforcement learning (ICRL), focusing on the Decision-Pretrained Transformer (DPT, Lee et al., 2023). To address the challenge of reward poisoning attacks targeting the DPT, we propose a novel adversarial training framework, called Adversarially Trained DPT (AT-DPT). Our method simultaneously trains a population of attackers to minimize the true reward of the DPT by poisoning environment rewards, and a DPT model to infer optimal actions from the poisoned data. We evaluate the effectiveness of our approach against standard bandit algorithms, including robust baselines designed to handle reward contamination. Our results show that AT-DPT significantly outperforms them in bandit settings under a learned attacker, and generalizes to more complex environments such as adaptive attackers and MDPs. It shows promise in ICRL as a meta-RL approach to learning effective corruption-robust algorithms.

Deep Learning · Large Language Models

Nikhil Chandak, Shashwat Goel, Ameya Pandurang Prabhu, Moritz Hardt, Jonas Geiping

High-stakes decision making involves reasoning under uncertainty about the future. In this work, we train language models to make predictions on open-ended forecasting questions. To scale up training data, we synthesize novel forecasting questions from global events reported in daily news. While directly training on this data leads to performance drops, carefully curating questions creates a valuable training resource. We use the resulting dataset, OpenForesight, to post-train Qwen3 thinking models. To prevent leakage of future information during training and evaluation, we use an offline news corpus, both for data generation and retrieval in our forecasting system. Guided by a small validation set, we show the benefits of retrieval, and an improved reward function for reinforcement learning (RL). Once we obtain our final forecasting system, we perform held-out testing between May to August 2025. Our specialized model, OpenForecaster-8B, matches much larger proprietary models, with our training improving the accuracy, calibration, and consistency of predictions. We find calibration improvements from forecasting training generalize across popular benchmarks. We will open-source our models, code, and data to make LLM based forecasting research broadly accessible.

Deep Learning · Large Language Models

Yasi Zhang, Tianyu Chen, Mingyuan Zhou, Oscar Leong, Ying Nian Wu, Michal Lukasik

Large language models (LLMs) are increasingly deployed as automated evaluators that assign numeric scores to model outputs, a paradigm known as LLM-as-a-Judge. However, standard Reinforcement Learning (RL) methods typically rely on binary rewards (e.g., 0-1 accuracy), thereby ignoring the ordinal structure inherent in regression tasks; for instance, they fail to recognize that predicting 4 is significantly better than predicting 1 when the ground truth is 5. Conversely, existing regression-aware approaches are often confined to Supervised Fine-Tuning (SFT), limiting their ability to explore optimal reasoning paths. To bridge this gap, we propose REAL (Regression-Aware Reinforcement Learning), a principled RL framework designed to optimize regression rewards, and also proven to be optimal for correlation metrics. A key technical challenge is that the regression objective is explicitly policy-dependent, thus invalidating standard policy gradient methods. To address this, we employ the generalized policy gradient estimator, which naturally decomposes optimization into two complementary components: (1) exploration over Chain-of-Thought (CoT) trajectory, and (2) regression-aware prediction refinement of the final score. Extensive experiments across model scales (8B to 32B) demonstrate that REAL consistently outperforms both regression-aware SFT baselines and standard RL methods, exhibiting significantly better generalization on out-of-domain benchmarks. On Qwen3-32B specifically, we achieve gains of +8.40 Pearson and +7.20 Spearman correlation over the SFT baseline, and +18.30/+11.20 over the base model. These findings highlight the critical value of integrating regression objectives into RL exploration for accurate LLM evaluation.

Deep Learning · Large Language Models

Johannes Ackermann, Michael Noukhovitch, Takashi Ishida, Masashi Sugiyama

Reinforcement Learning from Human Feedback (RLFH) or Verifiable Rewards (RLVR) are two key steps in the post-training of modern Language Models (LMs). A common problem is reward hacking, where the policy may exploit inaccuracies of the reward and learn an unintended behavior. Most previous works address this by limiting the policy update with a Kullback-Leibler (KL) penalty towards a reference model. We propose a different framing: Train the LM in a way that biases policy updates towards regions in which the reward is more accurate. First, we derive a theoretical connection between the accuracy of a reward model and the flatness of an optimum at convergence. Gradient regularization (GR) can then be used to bias training to flatter regions and thereby maintain reward model accuracy. We confirm these results by showing that the gradient norm and reward accuracy are empirically correlated in RLHF. We then show that Reference Resets of the KL penalty implicitly use GR to find flatter regions with higher reward accuracy. We further improve on this by proposing to use explicit GR with an efficient finite-difference estimate. Empirically, GR performs better than a KL penalty across a diverse set of RL experiments with LMs. GR achieves a higher GPT-judged win-rate in RLHF, avoids overly focusing on the format in rule-based math rewards, and prevents hacking the judge in LLM-as-a-Judge math tasks.

Deep Learning · Large Language Models

Utsav Singh, Sidhaarth Murali, Souradip Chakraborty, Amrit Singh Bedi

Reinforcement learning with verifiable rewards can improve LLM reasoning, but learning is sample-inefficient under sparse terminal rewards. Prior work mitigates this by adding natural language critiques, yet it typically treats critique generation as fixed or auxiliary, so correct-sounding feedback may not translate into higher verified reward. We argue that natural language actor-critic for reasoning is inherently bilevel: the usefulness of the critique is defined by its downstream effect on the actor after adaptation. We formalize this coupling as a Stackelberg bilevel program and derive Bilevel Natural Language Actor-Critic (Bi-NAC), which jointly trains a critic to generate reward-improving feedback and an actor to exploit it. Across reasoning benchmarks, Bi-NAC improves sample and parameter efficiency over RL baselines and fixed-critic feedback methods. We perform experiments on MATH-500, MBPP, and GPQA demonstrating that Bi-NAC significantly enhances parameter and sample efficiency, enabling smaller models to outperform larger baselines. Specifically, our 2B model consistently outperforms the larger 3B GRPO baseline across all tasks (e.g., 46.6% vs. 41.4% on MATH-500), while our 6B model surpasses the 7B GRPO baseline (e.g., 49.3% vs. 43.6% on GPQA). These results show that aligning actor and critic via bilevel formulation provides a robust and efficient alternative for solving complex reasoning tasks.

Theory · Reinforcement Learning and Planning

Jose Aguilar Escamilla, Haoyang Hong, Jiawei Li, Haoyu Zhao, Xuezhou Zhang, Sanghyun Hong, Huazheng Wang

We study reward poisoning attacks in reinforcement learning (RL), where an adversary manipulates rewards within constrained budgets to force the target RL agent to adopt a policy that aligns with the attacker's objectives. Prior works on reward poisoning mainly focused on sufficient conditions to design a successful attacker, while only a few studies discussed the infeasibility of targeted attacks. This paper provides the first precise necessity and sufficiency characterization of the attackability of a linear MDP under reward poisoning attacks. Our characterization draws a bright line between the vulnerable RL instances, and the intrinsically robust ones which cannot be attacked without large costs even running vanilla non-robust RL algorithms. Our theory extends beyond linear MDPs---by approximating deep RL environments as linear MDPs, we show that our theoretical framework effectively distinguishes the attackability and efficiently attacks the vulnerable ones, demonstrating both the theoretical and practical significance of our characterization.

Reinforcement Learning · Deep RL

Chengxuan Lu, Zhenquan Zhang, Shukuan Wang, Qunzhi Lin, Baigui Sun, Yang Liu

Post-training with reinforcement learning (RL) has recently shown strong promise for advancing multimodal agents beyond supervised imitation. However, RL remains limited by poor data efficiency, particularly in settings where interaction data are scarce and quickly become outdated. To address this challenge, GIPO (Gaussian Importance sampling Policy Optimization) is proposed as a policy optimization objective based on truncated importance sampling, replacing hard clipping with a log-ratio–based Gaussian trust weight to softly damp extreme importance ratios while maintaining non-zero gradients. Theoretical analysis shows that GIPO introduces an implicit, tunable constraint on the update magnitude, while concentration bounds guarantee robustness and stability under finite-sample estimation. Experimental results show that GIPO achieves state-of-the-art performance among clipping-based baselines across a wide range of replay buffer sizes, from near on-policy to highly stale data, while exhibiting superior bias–variance trade-off, high training stability and improved sample efficiency. Code is provided in supplementary material.

Deep Learning · Large Language Models

Zhaoxi Zhang, Yitong Duan, Yanzhi Zhang, Yiming Xu, Zhixiang Wang, Kun Liang, Yang Li, Jiahui Liang, Deguo Xia, Jizhou Huang 等

Locating files and functions requiring modification in large software repositories is challenging due to their scale and structural complexity. Existing LLM-based methods typically treat this as a repository-level retrieval task and rely on multiple auxiliary tools, which often overlook code execution logic and complicate model control. We propose RepoNavigator, an LLM agent equipped with a single execution-aware tool: jumping to the definition of an invoked symbol. This unified design reflects the actual flow of code execution while simplifying tool manipulation. RepoNavigator is trained end-to-end via Reinforcement Learning (RL) directly from a base pretrained model, without relying on closed-source distillation. Experiments demonstrate that RL-trained RepoNavigator achieves state-of-the-art performance, with the 7B model outperforming 14B baselines, the 14B model surpassing 32B competitors, and the 32B model exceeding closed-source models such as GPT-5 on most metrics. These results confirm that integrating a single, structurally grounded tool with RL training provides an efficient and scalable solution for repository-level issue localization.

Reinforcement Learning · Deep RL

Lipeng Zu, YU QIAN, Shayok Chakraborty, Xiaonan Zhang

Offline-to-online reinforcement learning (O2O RL) faces a central challenge between retaining offline conservatism and adapting to online feedback under distribution shift. This challenge arises because data behavior evolves during fine-tuning, rendering data origin a misleading basis for constraint handling and thereby leading to objective–data mismatch. We therefore propose Dynamic Alignment for RElease (DARE), a distribution-aware framework for sample-level constraint release based on the behavioral consistency with a behavior model. To our knowledge, DARE is the first to condition constraint release on behavioral consistency via a posterior-induced exchange mechanism, moving beyond a binary offline/online data distinction. Importantly, DARE requires only per-sample behavioral alignment, enabling instantiation on top of many offline algorithms with flexible choices of behavior models and fine-tuning objectives. We provide a theoretical analysis showing that behavior-based sample exchange consistently improves the distinction between offline-like and online-like subsets. Experiments on D4RL demonstrate that DARE consistently improves fine-tuning stability and achieves superior final performance over strong offline-to-online baselines. (Anonymized code archive is included in the supplementary material.)

Deep Learning · Generative Models and Autoencoders

Jili Chen, Changqin Huang, Qionghao Huang, Yaxin Tu, Zhonglong Zheng, Xiaodi Huang

Flow-matching has emerged as a leading framework for high-fidelity text-to-image generation. However, its alignment with human preferences through RL is often hindered by substantial computational overhead. In this paper, we introduce Flow-TTRL, the first test-time reinforcement learning framework that achieves alignment on the fly. Our approach reinterprets intermediate latent representations as an implicit policy and utilizes SDE-based rollouts to explore high-reward trajectories within the learned vector field. Specifically, we propose a two-stage optimization strategy: Proximal Reward Difference Prediction (PRDP) ensures structural stability in high-noise regimes through pairwise reward regression, while Group Relative Policy Optimization (GRPO) refines fine-grained aesthetic details by maximizing relative advantages within sampled candidate groups. Experimental results show that Flow-TTRL significantly boosts aesthetic quality, text-image alignment, and human preference across diverse backbones. On the GenEval benchmark, Flow-TTRL elevates the accuracy of SD 3.5-Medium from 63\% to 87\% and Flux.1 Dev from 66\% to 83\%. Furthermore, our framework achieves an average gain of 15\% to 20\% across T2I-CompBench metrics, delivering performance comparable to state-of-the-art RL-based fine-tuning methods without the need for additional fine-tuning.

Deep Learning · Large Language Models

Yapei Chang, Kyle Lo, Mohit Iyyer, Luca Soldaini

Generating step-by-step "how-to" procedures is a key LLM capability: how-to advice is commonly requested in chatbots, and step-by-step planning is critical for reasoning over complex tasks. Yet, measuring and improving procedural validity at scale on real-world tasks remains challenging and understudied. To address this, we introduce How2Everything, a scalable framework to evaluate and improve goal-conditioned procedure generation. Our pipeline How2Mine extracts and rewrites 351K procedures from 980K web pages across 14 topics, and can scale to larger corpora. From this pool we build How2Bench, a 7K-example evaluation set balanced across topics. We also introduce How2Score, an evaluation protocol that uses an LLM judge to detect whether a generation contains any critical failure that would prevent achieving the goal. For low-cost, reproducible evaluation, we distill a frontier judge into an open 8B model achieving 80.5\% agreement with human annotators. How2Bench reveals clear scaling trends across model size and training stages, providing signal early in pretraining. Finally, RL using How2Score as a reward improves performance on How2Bench by >10 points across three base models without systematic regressions on standard benchmarks, with gains not primarily explained by source-document memorization or superficial format compliance. We release all code and data upon acceptance.

Deep Learning · Large Language Models

Zhepei Wei, Xiao Yang, Kai Sun, Jiaqi Wang, Rulin Shao, Jingxiang Chen, Mohammad Kachuee, Teja Gollapudi, Yiwei Liao, Nicolas SCHEFFER 等

While large language models (LLMs) have demonstrated strong performance on factoid question answering, they are still prone to hallucination and untruthful responses, particularly when tasks demand information outside their parametric knowledge. Indeed, truthfulness requires more than accuracy---models must also recognize uncertainty and abstain when unsure to avoid hallucinations. This presents a fundamental challenge for existing methods: approaches that optimize for accuracy often amplify hallucinations, while those that encourage abstention can become overly conservative, sacrificing correct answers. Both extremes ultimately compromise truthfulness. In this work, we present TruthRL, a general reinforcement learning (RL) framework that directly optimizes the truthfulness of LLMs. Specifically, we implement TruthRL using GRPO with a simple yet effective ternary reward that distinguishes correct answers, hallucinations, and abstentions. It incentivizes models to reduce hallucinations not only by providing correct responses, but also by enabling abstention when uncertain, thereby improving truthfulness. Extensive experiments across four knowledge-intensive benchmarks show that \model significantly reduces hallucinations (e.g., 43.5\% $\rightarrow$ 19.4\%) and improves truthfulness (e.g., 5.3\% $\rightarrow$ 37.2\%), with consistent gains across various backbone models (e.g., Qwen, Llama). In-depth ablation study demonstrates that vanilla accuracy-driven methods such as supervised fine-tuning or RL with a binary reward struggle to balance factual correctness and uncertainty, whereas the truthfulness-driven TruthRL achieves strong performance in both accuracy and truthfulness, underscoring the importance of learning objective design for developing truthful LLMs. Moreover, we find the improvement of \model arises from enhancing the capability of LLMs to recognize their knowledge boundary, hence avoiding being overly conservative as the baselines are. Further analysis validates our method across multiple evaluation judges, and confirms that TruthRL is robust to hallucination-baiting questions.

Deep Learning · Large Language Models

Fahim Tajwar, Guanning Zeng, Yueer Zhou, Yuda Song, Daman Arora, Yiding Jiang, Jeff Schneider, Russ Salakhutdinov, Haiwen Feng, Andrea Zanette

Maximum likelihood is fundamental to supervised learning but it cannot be directly applied in correctness-based problems with non-differentiable sampling. In these settings, reinforcement learning (RL) is typically used to maximize expected reward. We show that for binary correctness tasks, expected-reward RL is a first-order approximation of the maximum likelihood objective, yielding vanishing learning signal on low-success inputs. We introduce **Maximum Likelihood Reinforcement Learning (MaxRL)**, a compute-indexed family of sampling-based objectives derived from a pass@k expansion of the likelihood, which interpolates between standard RL and exact maximum likelihood as compute increases. MaxRL admits a simple unbiased policy-gradient estimator whose optimized objective improves with additional compute. Across multiple domains, MaxRL consistently outperforms standard RL and GRPO, achieving higher $pass@1$ and substantially improved $pass@k$.

Reinforcement Learning · Deep RL

Mahsa Bastankhah, Sophie Broderick, Benjamin Eysenbach

In many practical reinforcement learning (RL) environments, observations are far higher-dimensional than the variables that matter for control. In this work, we ask: can we learn representations that capture only control-relevant features of the environment? We study this question through the \emph{empowerment} objective, which maximizes an agent’s influence over the environment and is widely used for unsupervised skill learning. We show that empowerment agents induce two distinct representations --- forward and backward --- that capture complementary aspects of the state, and both of which are invariant to control-irrelevant features. Thus, empowerment maximization leads agents to learn an implicit, \emph{control-centric} model of the world. Our analysis highlights the importance of learning representations through interaction rather than from passive datasets: interaction aimed at maximizing control is essential for learning useful invariance properties, a perspective that aligns closely with the causal learning literature.

Reinforcement Learning · Deep RL

Jiafei Lyu, Zichuan Lin, Scott Fujimoto, Kai Yang, Yangkun Chen, Saiyong Yang, Zongqing Lu, Deheng Ye

Model-based representations recently stand out as a promising framework that embeds latent dynamics information into the representations for downstream off-policy actor-critic learning. It implicitly combines the advantages of both model-free and model-based approaches while avoiding the training costs associated with model-based methods. Nevertheless, existing model-based representation methods can fail to capture sufficient information about relevant variables and can overfit to early experiences in the replay buffer. These incur biases in representation and actor-critic learning, leading to inferior performance. To address this, we propose Debiased model-based Representations for Q-learning, tagged DR.Q algorithm. DR.Q explicitly maximizes the mutual information between the representations of the current state-action pair and the next state besides minimizing their deviations, and samples transitions with faded prioritized experience replay. We evaluate DR.Q on numerous continuous control benchmarks with a single set of hyperparameters, and the results demonstrate that DR.Q can match or surpass recent strong baselines, sometimes outperforming them by a large margin.

Deep Learning · Large Language Models

Amrith Setlur, Zijian Wang, Andrew Cohen, Paria Rashidinejad, Sang Michael Xie

Typical reinforcement learning (RL) methods for LLM reasoning waste compute on hard problems, where correct on-policy traces are rare and policy gradients vanish. To bootstrap more efficient RL, we consider reusing old sampling FLOPs (from prior inference or RL training) in the form of off-policy traces. We introduce PrefixRL, where we condition on the prefix of successful off-policy traces and run on-policy RL to complete them, side-stepping instabilities from using off-policy data as supervision targets. PrefixRL boosts the learning signal on hard problems by modulating the difficulty of the problem through the off-policy prefix length. We prove that the PrefixRL objective is not only consistent with the standard RL objective but also more sample efficient. Empirically, we discover back-generalization: training only on prefixed problems generalizes to out-of-distribution unprefixed performance, with learned strategies often differing from those in the prefix. In our experiments, we source the off-policy traces by rejection sampling with the base model, creating a self-improvement loop. On hard reasoning problems, PrefixRL reaches the same training reward 2x faster than the strongest baseline (SFT on off-policy data then RL), even after accounting for the compute spent on the initial rejection sampling, and increases the final reward by 3x.

Deep Learning · Large Language Models

Xilun Chen, Ilia Kulikov, Vincent-Pierre Berges, Barlas Oğuz, Rulin Shao, Gargi Ghosh, Scott Yih

Reasoning Large Language Models (R-LLMs) have significantly advanced complex reasoning tasks but often struggle with factuality, generating substantially more hallucinations than their non-reasoning counterparts on long-form factuality benchmarks. However, extending online Reinforcement Learning (RL), a key component in recent R-LLM advancements, to the long-form factuality setting poses several unique challenges due to the lack of reliable verification methods. Previous work has utilized automatic factuality evaluation frameworks such as FActScore to curate preference data in the offline RL setting, yet we find that directly leveraging such methods as the reward in online RL leads to reward hacking in multiple ways, such as producing less detailed or relevant responses. We propose a novel reward function that simultaneously considers the factual precision, response detail level, and answer relevance, and applies online RL to learn high quality factual reasoning. Evaluated on six long-form factuality benchmarks, our factual reasoning model achieves an average reduction of 23.1 percentage points in hallucination rate, a 23% increase in answer detail level, and no degradation in the overall response helpfulness.

Deep Learning · Large Language Models

Jingyuan Yan, Qingchen Liu, Qichao Ma, Jiahu Qin

Reinforcement Learning (RL) with Group Relative Policy Optimization (GRPO) shows great promise for enhancing LLM reasoning, but remains challenged by sparse and unstable rewards in long-horizon tasks. Existing approaches to reward shaping struggle to balance semantic expressiveness, reliability, and computational efficiency: heuristic rules lack flexibility, while LLM-as-a-Judge incurs high computational cost and suffer from inconsistent and misaligned scoring signals in long-context settings. To address these challenges, we introduce GLARE, a neuro-symbolic reward framework that decouples semantic abstraction from credit assignment. Specifically, to leverage semantic understanding while preserving symbolic determinism, we first extract and symbolize trajectory events into a discrete representation. These events are then translated into Linear Temporal Logic (LTL) formulas, which are compiled into deterministic automata that track the agent's progress via state transitions. This mechanism yields dense and consistent reward signals, avoiding unstable direct scoring while significantly reducing computational cost. Empirical results on ALFWorld show that GLARE outperforms GRPO by 12.1\% in success rate, while achieving an 8.1\% improvement over conventional LLM-based judges using only 15\% of their computational cost.

Reinforcement Learning · Deep RL

Bhavya Agrawalla, Michal Nauman, Aviral Kumar

Recent work shows that flow-matching networks can be effective for value function estimation in reinforcement learning, but it remains unclear why they work well or whether flow-matching Q-functions differ fundamentally from standard critics. We show that their success is not explained by distributional RL: explicitly modeling return distributions often degrades performance. Instead, we argue that flow-matching Q-functions are effective because they couple a learned velocity field with an integration procedure that is used both during training and to read out Q-values at inference time. This coupling enables robust value prediction through \emph{test-time recovery} from imperfect intermediate estimates where errors dampen out as more integration steps are performed. This mechanism is absent in monolithic critics. Beyond test-time recovery, training with the integration procedure induces more \emph{plastic} representations, allowing critics to represent non-stationary future TD targets without overwriting previous features. We formalize these effects and validate them empirically, showing that flow-matching critics outperform monolithic critics by over $2\times$ in performance and achieve $5$–$10\times$ higher sample efficiency in high-UTD regimes.