Yueyang Zhang

Chao Huang, Yujing Lu, Quangang Li et al.

Entropy regularization is a standard technique in reinforcement learning (RL) to enhance exploration, yet it yields negligible effects or even degrades performance in Large Language Models (LLMs). We attribute this failure to the cumulative tail risk inherent to LLMs with massive vocabularies and long generation horizons. In such environments, standard global entropy maximization indiscriminately dilutes probability mass into the vast tail of invalid tokens rather than focusing on plausible candidates, thereby disrupting coherent reasoning. To address this, we propose Trust Region Entropy (TRE), a method that encourages exploration strictly within the model's trust region. Extensive experiments across mathematical reasoning (MATH), combinatorial search (Countdown), and preference alignment (HH) tasks demonstrate that TRE consistently outperforms vanilla PPO, standard entropy regularization, and other exploration baselines. Our code is available at https://github.com/WhyChaos/TRE-Encouraging-Exploration-in-the-Trust-Region.

Yu Liang, Liangxin Liu, Longzheng Wang et al.

Generative reward models (GRMs) have emerged as a promising approach for aligning Large Language Models (LLMs) with human preferences by offering greater representational capacity and flexibility than traditional scalar reward models. However, GRMs face two major challenges: reliance on costly human-annotated data restricts scalability, and self-training approaches often suffer from instability and vulnerability to reward hacking. To address these issues, we propose ConsistRM, a self-training framework that enables effective and stable GRM training without human annotations. ConsistRM incorporates the Consistency-Aware Answer Reward, which produces reliable pseudo-labels with temporal consistency, thereby providing more stable model optimization. Moreover, the Consistency-Aware Critique Reward is introduced to assess semantic consistency across multiple critiques and allocates fine-grained and differentiated rewards. Experiments on five benchmark datasets across four base models demonstrate that ConsistRM outperforms vanilla Reinforcement Fine-Tuning (RFT) by an average of 1.5%. Further analysis shows that ConsistRM enhances output consistency and mitigates position bias caused by input order, highlighting the effectiveness of consistency-aware rewards in improving GRMs.

Kai Qin, Liangxin Liu, Yu Liang et al.

Reward Models (RMs) are critical components in the Reinforcement Learning from Human Feedback (RLHF) pipeline, directly determining the alignment quality of Large Language Models (LLMs). Recently, Generative Reward Models (GRMs) have emerged as a superior paradigm, offering higher interpretability and stronger generalization than traditional scalar RMs. However, existing methods for GRMs focus primarily on outcome-level supervision, neglecting analytical process quality, which constrains their potential. To address this, we propose ReflectRM, a novel GRM that leverages self-reflection to assess analytical quality and enhance preference modeling. ReflectRM is trained under a unified generative framework for joint modeling of response preference and analysis preference. During inference, we use its self-reflection capability to identify the most reliable analysis, from which the final preference prediction is derived. Experiments across four benchmarks show that ReflectRM consistently improves performance, achieving an average accuracy gain of +3.7 on Qwen3-4B. Further experiments confirm that response preference and analysis preference are mutually reinforcing. Notably, ReflectRM substantially mitigates positional bias, yielding +10.2 improvement compared with leading GRMs and establishing itself as a more stable evaluator.

Min Cai, Yu Liang, Longzheng Wang et al.

Reinforcement learning (RL) has played a central role in recent advances in large reasoning models (LRMs), yielding strong gains in verifiable and open-ended reasoning. However, training a single general-purpose LRM across diverse domains remains challenging due to pronounced domain heterogeneity. Through a systematic study of two widely used strategies, Sequential RL and Mixed RL, we find that both incur substantial cross-domain interference at the behavioral and gradient levels, resulting in limited overall gains. To address these challenges, we introduce **M**odular **G**radient **S**urgery (**MGS**), which resolves gradient conflicts at the module level within the transformer. When applied to Llama and Qwen models, MGS achieves average improvements of 4.3 (16.6\%) and 4.5 (11.1\%) points, respectively, over standard multi-task RL across three representative domains (math, general chat, and instruction following). Further analysis demonstrates that MGS remains effective under prolonged training. Overall, our study clarifies the sources of interference in multi-domain RL and presents an effective solution for training general-purpose LRMs.

Ruishan Guo, Yibing Liu, Guoxin Ma et al.

Scaling up model parameters has long been a prevalent training paradigm driven by the assumption that larger models yield superior generation capabilities. However, under lossy context compression in a compressor-decoder setup, we observe a Size-Fidelity Paradox: increasing the compressor size can lessen the faithfulness of reconstructed contexts though training loss decreases. Through extensive experiments across models from 0.6B to 90B, we coin this paradox arising from two dominant factors: 1) knowledge overwriting: larger models increasingly replace source facts with their own prior beliefs, e.g., ``the white strawberry'' $\to$ ``the red strawberry''; and 2) semantic drift: larger models tend to paraphrase or restructure content instead of reproducing it verbatim, e.g., ``Alice hit Bob'' $\to$ ``Bob hit Alice''. By holding model size fixed, we reflect on the emergent properties of compressed context representations. We show that the culprit is not parameter count itself, but the excessive semantic capacity and amplified generative uncertainty that accompany scaling. Specifically, the increased rank of context embeddings facilitates prior knowledge intrusion, whereas higher entropy over token prediction distributions promotes rewriting. Our results complement existing evaluations over context compression paradigm, underpinning a breakdown in scaling laws for faithful preservation in open-ended generation.