Difficulty-Estimated Policy Optimization (DEPO) is a new framework that improves the efficiency and robustness of reasoning alignment in Large Reasoning Models. It introduces an online Difficulty Estimator that filters training data before rollouts, ensuring computational resources focus on high‑learning‑potential samples. Empirical results show DEPO achieves about 1.86× faster rollouts than DAPO and outperforms GRPO on benchmarks at similar training costs.
By Yu Zhao, Fan Jiang, Tianle Liu, Bo Zeng, Xintong Wang, Jiaxuan Teng, Yu Liu, Yichao Du
The paper introduces Geometric Anchor Preference Optimization (GAPO), a method that replaces the static reference policy in Direct Preference Optimization with a dynamic, geometry-aware anchor—a small adversarial perturbation of the current policy. GAPO uses this anchor to adaptively reweight preference pairs based on local sensitivity, and defines an Anchor Gap that approximates worst‑case local margin degradation. Experiments show that GAPO improves robustness to noisy supervision while matching or surpassing existing LLM alignment and reasoning benchmarks.
By Youngjae Cho, Jongsuk Kim, Ji-Hoon Kim
The paper introduces Gradient-Aligned Reward (GAR), a reinforcement learning technique that uses truncated backpropagation to generate a compact gradient vector for each rollout and compares it to an expert-anchor gradient via cosine similarity. This dense, reasoning-aware reward improves large language model chain-of-thought reasoning on math benchmarks and transfers to other tasks without domain‑specific data, while adding less than 9% computational overhead. GAR outperforms existing baselines such as GRPO on Qwen3-4B and Qwen3-8B models.
By Leqi Zheng, Jinbo Su, Fang Niu, Chaokun Wang, Weiping Wang, Jiajun Zhang, Shannan Yan, Jie Wu, Zhaolu Kang, Rong Fu, Hang Zhang
arXiv:2606. 04807v1 Announce Type: new Abstract: Mitigating social bias in Large Language Models (LLMs) presents a distinct alignment challenge: unlike verifiable tasks, bias lacks a single ground truth, creating a high-variance, subjective reward landscape.
By Saket Reddy, Ke Yang, ChengXiang Zhai
arXiv:2606. 01281v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for enhancing the reasoning capabilities of large language models (LLMs).
By Yixiu Mao, Yun Qu, Qi Wang, Heming Zou, Xiangyang Ji
arXiv:2603. 25184v2 Announce Type: replace-cross Abstract: Reinforcement learning (RL) has become essential for post-training large language models (LLMs) in reasoning tasks.
By Jiahao Wu, Ning Lu, Shengcai Liu, Kun Wang, Yanting Yang, Bailong Lin, Chen Jason Zhang, Li Qing, Ke Tang
The paper addresses instability in off‑policy reinforcement learning for large language models (LLMs) caused by large policy lag. It shows that the correct correction for the prefix‑marginal policy gradient is the cumulative prefix importance ratio, which token‑level weighting fails to capture. To improve stability, the authors propose Minimum Prefix Ratio (MinPRO), a surrogate objective that uses the minimum token‑level ratio in the preceding prefix, and demonstrate its effectiveness on dense and mixture‑of‑experts LLMs across mathematical reasoning benchmarks.
By Shiye Lei, Zhihao Cheng, Dacheng Tao
Mitigating social bias in Large Language Models (LLMs) presents a distinct alignment challenge: unlike verifiable tasks, bias lacks a single ground truth, creating a high-variance, subjective reward landscape. Previous preference-based fine-tuning methods have major trade-offs: Direct Preference Optimization (DPO) is limited by the lack of exploration inherent in offline training, while Proximal Policy Optimization (PPO) can lead to training instability due to potentially unreliable critic estimates.
arXiv:2608. 05541v1 Announce Type: new Abstract: Evolution Strategy (ES) is a promising alternative to gradient-based fine-tuning for resource-constrained Large Language Model (LLM) reasoning.
By Yu Gu, Zhi Zheng, Yunpeng Ba, Xialiang Tong, Mingxuan Yuan, Zhenkun Wang
RollVerify is a lightweight reinforcement learning framework that addresses the trade‑off between efficiency and accuracy in long‑tail rollout settings. It introduces an off‑policy shift metric (OPS) to quantify deviation in partially generated trajectories and uses sequence‑level and token‑level verification to truncate invalid suffixes before training. Experiments on mathematical reasoning tasks show that RollVerify matches on‑policy accuracy while cutting training costs, with preliminary evidence from code‑generation tasks.
By Yongqiang Yao, Jinru Tan, Kaihuan Liang, Zixin Yin, Yazhe Niu, Ruihao Gong, Dahua Lin, Ningyi Xu
Continuous-Utility Direct Preference Optimization (CU‑DPO) replaces binary preference labels with continuous scores to better capture fine‑grained reasoning quality. The framework aligns language models to multiple prompt‑based cognitive strategies, proving a θ(K log K) sample‑complexity improvement and convergence to an entropy‑regularized utility‑maximizing policy. A two‑stage pipeline—strategy selection via best‑vs‑all comparisons and execution refinement using margin‑stratified pairs—yields significant gains on mathematical reasoning, code generation, and causal reasoning benchmarks.
By Muhammad Ahmed Mohsin, Muhammad Umer, Emily Fox
arXiv:2606. 02521v1 Announce Type: new Abstract: One-step text-to-image generators are attractive for deployment because they generate an image with a single forward pass, but preference finetuning them remains difficult: standard alignment methods often rely on policy likelihoods, denoising trajectories, differentiable reward gradients, or test-time optimization.
By Zhou Jiang, Yandong Wen, Zhen Liu