Significant progress in reward modeling over recent years has been driven by a paradigm shift from task-specific designs towards generalist reward models. Despite this trend, developing effective reward models remains a fundamental challenge: the heavy reliance on large-scale labeled preference data. Pre-training on abundant unlabeled data offers a promising direction, but existing approaches fall short of instilling explicit reasoning into reward models. To bridge this gap, we propose a self-training approach that leverages unlabeled data to elicit reward reasoning in reward models. Based on this approach, we develop GRAM-R$^2$, a generative reward model trained to produce not only preference labels but also accompanying reward rationales. GRAM-R$^2$ can serve as a foundation model for reward reasoning and can be applied to a wide range of tasks with minimal or no additional fine-tuning. It can support downstream applications such as response ranking and task-specific reward tuning. Experiments on response ranking, task adaptation, and reinforcement learning from human feedback demonstrate that GRAM-R$^2$ consistently delivers strong performance, outperforming several strong discriminative and generative baselines.

We study whether self-learning can scale LLM-based agents without relying on human-curated datasets or predefined rule-based rewards. Through controlled experiments in a search-agent setting, we identify two key determinants of scalable agent training: the source of reward signals and the scale of agent task data. We find that rewards from a Generative Reward Model (GRM) outperform rigid rule-based signals for open-domain learning, and that co-evolving the GRM with the policy further boosts performance. Increasing the volume of agent task data-even when synthetically generated-substantially enhances agentic capabilities. Building on these insights, we propose \textbf{Agentic Self-Learning} (ASL), a fully closed-loop, multi-role reinforcement learning framework that unifies task generation, policy execution, and evaluation within a shared tool environment and LLM backbone. ASL coordinates a Prompt Generator, a Policy Model, and a Generative Reward Model to form a virtuous cycle of harder task setting, sharper verification, and stronger solving. Empirically, ASL delivers steady, round-over-round gains, surpasses strong RLVR baselines (e.g., Search-R1) that plateau or degrade, and continues improving under zero-labeled-data conditions, indicating superior sample efficiency and robustness. We further show that GRM verification capacity is the main bottleneck: if frozen, it induces reward hacking and stalls progress; continual GRM training on the evolving data distribution mitigates this, and a small late-stage injection of real verification data raises the performance ceiling. This work establishes reward source and data scale as critical levers for open-domain agent learning and demonstrates the efficacy of multi-role co-evolution for scalable, self-improving agents. The data and code of this paper are released at https://github.com/forangel2014/Towards-Agentic-Self-Learning

Large language models (LLMs) are increasingly trusted as automated judges, assisting evaluation and providing reward signals for training other models, particularly in reference-based settings like Reinforcement Learning with Verifiable Rewards (RLVR). However, we uncover a critical vulnerability even in this reference-based paradigm: generative reward models are systematically susceptible to reward hacking. We find that superficial inputs, which we term ''master keys'' such as non-word symbols (e.g., '':'' or ''.'') or generic reasoning openers (e.g., ''Thought process:'' or ''Let's solve this problem step by step.''), can consistently elicit false positive rewards without any substantive reasoning. Our systematic evaluation demonstrates this is a widespread failure affecting a diverse range of models, including leading proprietary systems such as GPT-o1 and Claude-4. These results challenge the assumed robustness of LLM judges and pose a significant threat to their reliability. To address this, we propose a simple yet effective data augmentation strategy using truncated model outputs as adversarial negative examples. The resulting Master Reward Models (Master-RMs) demonstrate state-of-the-art robustness against these ''master key'' attacks while maintaining high performance in standard evaluation settings. We supplement these findings with a comprehensive analysis of the vulnerability across model scales, prompt variations, and common inference-time strategies, offering insights to guide future research on robust LLM evaluation. We release our robust, general-domain reward models and the synthetic training data at https://huggingface.co/sarosavo/Master-RM and https://huggingface.co/datasets/sarosavo/Master-RM.

Reinforcement Learning from Human Feedback (RLHF) has emerged as a prominent paradigm for training large language models and multimodal systems. Despite the notable advances enabled by existing RLHF training frameworks, significant challenges remain to scale to complex multimodal workflows and adapt to dynamic workloads. In particular, current systems often encounter limitations related to controller scalability when managing large models, as well as inefficiencies in orchestrating intricate RLHF pipelines, especially in scenarios that require dynamic sampling and resource allocation. In this paper, we introduce WeChat-YATT Yet Another Transformer Trainer in WeChat, a simple, scalable, and balanced RLHF training framework specifically designed to address these challenges. WeChat-YATT features a parallel controller programming model that enables flexible and efficient orchestration of complex RLHF workflows, effectively mitigating bottlenecks associated with centralized controller architectures and facilitating scalability in large-scale data scenarios. In addition, we propose a dynamic placement schema that adaptively partitions computational resources and schedules workloads, thereby significantly reducing hardware idle time and improving GPU utilization under variable training conditions. We evaluate WeChat-YATT across diverse experimental scenarios, demonstrating its substantial throughput improvements over state-of-the-art RLHF training frameworks. Furthermore, WeChat-YATT has been successfully deployed to train models that support WeChat product features for a large-scale user base, underscoring its effectiveness and robustness in real-world applications. We have made WeChat-YATT publicly available at https://www.github.com/tencent/WeChat-YATT.

Reinforcement learning (RL) has significantly improved the reasoning ability of large language models. However, current reward models underperform in challenging reasoning scenarios and predominant RL training paradigms rely on rule-based or reference-based rewards, which impose two critical limitations: 1) the dependence on finely annotated reference answer to attain rewards; and 2) the requirement for constrained output format. These limitations fundamentally hinder further RL data scaling and sustained enhancement of model reasoning performance. To address these limitations, we propose a comprehensive framework for evaluating and improving the performance of reward models in complex reasoning scenarios. We first present a reasoning-oriented benchmark (Libra Bench), systematically constructed from a diverse collection of challenging mathematical problems and advanced reasoning models, to address the limitations of existing reward model benchmarks in reasoning scenarios. We further introduce a novel approach for improving the generative reward model via learning-to-think methodologies. Based on the proposed approach, we develop Libra-RM series, a collection of generative reward models with reasoning capabilities that achieve state-of-the-art results on various benchmarks. Comprehensive downstream experiments are conducted and the experimental results demonstrate the correlation between our Libra Bench and downstream application, and the potential of Libra-RM to further improve reasoning models with unlabeled data.

Generative Reward Models (GRMs) provide greater flexibility than scalar reward models in capturing human preferences, but their effectiveness is limited by poor reasoning capabilities. This often results in incomplete or overly speculative reasoning paths, leading to hallucinations or missing key information in complex tasks. We address this challenge with ReasonGRM, a three-stage generative reward modeling framework. In the first stage, Zero-RL is used to generate concise, outcome-directed reasoning paths that reduce the likelihood of critical omissions. In the second stage, we introduce a novel evaluation metric, $R^\star$, which scores reasoning paths based on their generation likelihood. This favors paths that reach correct answers with minimal exploration, helping to reduce hallucination-prone data during training. In the final stage, the model is further refined through reinforcement learning on challenging examples to enhance its preference discrimination capabilities. Experiments on three public benchmarks show that ReasonGRM achieves competitive or state-of-the-art performance, outperforming previous best GRMs by 1.8\% on average and surpassing proprietary models such as GPT-4o by up to 5.6\%. These results demonstrate the effectiveness of reasoning-aware training and highlight the importance of high-quality rationale selection for reliable preference modeling.

Conversational recommendation systems (CRSs) use multi-turn interaction to capture user preferences and provide personalized recommendations. A fundamental challenge in CRSs lies in effectively understanding user preferences from conversations. User preferences can be multifaceted and complex, posing significant challenges for accurate recommendations even with access to abundant external knowledge. While interaction with users can clarify their true preferences, frequent user involvement can lead to a degraded user experience. To address this problem, we propose a generative reward model based simulated user, named GRSU, for automatic interaction with CRSs. The simulated user provides feedback to the items recommended by CRSs, enabling them to better capture intricate user preferences through multi-turn interaction. Inspired by generative reward models, we design two types of feedback actions for the simulated user: i.e., generative item scoring, which offers coarse-grained feedback, and attribute-based item critique, which provides fine-grained feedback. To ensure seamless integration, these feedback actions are unified into an instruction-based format, allowing the development of a unified simulated user via instruction tuning on synthesized data. With this simulated user, automatic multi-turn interaction with CRSs can be effectively conducted. Furthermore, to strike a balance between effectiveness and efficiency, we draw inspiration from the paradigm of reward-guided search in complex reasoning tasks and employ beam search for the interaction process. On top of this, we propose an efficient candidate ranking method to improve the recommendation results derived from interaction. Extensive experiments on public datasets demonstrate the effectiveness, efficiency, and transferability of our approach.

Reinforcement learning (RL) has been widely used in training large language models~(LLMs) for preventing unexpected outputs, \eg reducing harmfulness and errors. However, existing RL methods mostly adopt the instance-level reward, which is unable to provide fine-grained supervision for complex reasoning tasks, and can not focus on the few key tokens that lead to the incorrectness. To address it, we propose a new RL method named \textbf{RLMEC} that incorporates a generative model as the reward model, which is trained by the erroneous solution rewriting task under the minimum editing constraint, and can produce token-level rewards for RL training. Based on the generative reward model, we design the token-level RL objective for training and an imitation-based regularization for stabilizing RL process. And the both objectives focus on the learning of the key tokens for the erroneous solution, reducing the effect of other unimportant tokens. The experiment results on mathematical tasks and question-answering tasks have demonstrated the effectiveness of our approach. Our code and data are available at \url{https://github.com/RUCAIBox/RLMEC}.