grpo

Reviso is an agentic video captioning framework that uses hierarchical Context-Focus-Points planning and active perception to prioritize key visual evidence, trained with GRPO and specialized rewards, outperforming open-source baselines on four video captioning benchmarks.

The paper shows that a student model's traits can transfer purely through a judge's reward scores during GRPO post-training, even when outputs contain no semantic trace of the trait, and establishes a four-stage causal chain explaining when this subliminal transfer occurs, with implications for AI-safety dispositions.

CoFiBreak 提出一种粗到细的多样性奖励机制用于 GRPO 训练的多轮越狱攻击，在保持高效攻击的同时显著提升攻击策略多样性，并可迁移到未见过的前沿模型。

The paper proposes Li2SAO, a lightweight asynchronous RL method for LLM post-training that uses group-relative actor-critic bootstrapping to warm up the critic before switching to single-rollout GAE, catching up with or surpassing GRPO early in training and raising later performance ceilings.

A theory paper showing that in RLVR, every pass@1 improvement provably costs performance on some prompt under strategy specialization, with a tight bound on the number of pass@k crossover points, framed as an optimal tradeoff frontier rather than a bug.

SCOPE 提出选择性关键步骤的在线自蒸馏框架，通过错误定位与技能生成在 RL 训练中提供细粒度监督，在 ALFWorld 和 WebShop 上显著优于 GRPO 与 SEED。

The tweet shared Hugging Face’s tutorial “A curriculum for post-training agents,” which centers on coding agents and covers SFT, distillation, GRPO, OpenEnv multi-turn interactions, and training with a real harness, using experiments to clearly demonstrate the risks of reward hacking.

Anyscale is hosting a webinar on reinforcement learning for LLM post-training, covering RLVR with GRPO, on-policy distillation (OPD/MOPD), terminal agent post-training, and a deep dive into SkyRL's async RL capabilities.

SWE-Ladder 提出渐进式特权自蒸馏方法，将每个训练实例的验证修复转化为定位、复现、补丁三个阶段的专属证据，用于训练 SWE 智能体。在 SWE-bench Verified 上，Qwen3-4B 结合 R2E-Gym 框架的 pass@1 从 9.6% 提升至 16.4%，优于 GRPO 和 OPSD 基线。

The paper proposes ProfileSim, a profile-driven user simulation framework for financial marketing outbound calls, using GRPO with hierarchical rewards and jointly modeled user/business profiles to generate realistic, compliant dialogue data for task-oriented dialogue systems.

PhysEvolve 提出一个验证器驱动的自演化框架，利用物理引擎验证模型自生成的可执行世界与反事实仿真，以 GRPO 训练视觉语言模型的因果物理推理能力，在 PhysBench、CausalPhys 和 ContPhy 上取得显著提升且无需人类干预数据。

论文提出SATE-Bench交互式基准，包含316个真实场景与2525个多轮QA对，用于评估多模态大模型的视觉自我感知与主动全景视角探索能力。结果显示SOTA MLLM往往在证据不足时过早停止探索且自信度过高，而GRPO微调虽提升探索行为却因过度探索和多视角融合失败而降低整体准确率。

ReAPO 提出一种基于随机最短路径视角的重试感知策略优化方法，对 GRPO 进行改进，将额外生成分配给失败轨迹以实现更有原则的信用分配，在 Qwen3 和 OLMo-2 系列模型的数学与科学推理基准上显著提升首次尝试性能并减少生成 token 数。

The paper introduces GRSE (Group-Relative Skill Evolution), a GRPO-inspired method that assigns behavioral credit to agent–skill pairs in cooperative LLM teams by contrasting repeated executions, improving skill evolution on HotpotQA and LoCoMo over prior methods.

该论文提出将少量专家(SME)评判转化为密集奖励信号，并与检索指标结合成适用于GRPO的混合奖励，以解决非可验证领域（如金融服务检索）中稀疏奖励导致的梯度消失问题，实验显示Recall@3提升最多6%、MRR@3提升6.7%。

LongZero 让大语言模型自行合成多跳推理与多轮检索任务，并用 GRPO 在零外部数据的条件下自我提升长上下文推理能力，在五个长上下文基准上超越依赖外部语料的基线方法。

HarnessAdapt-RL jointly trains a coding policy and adapts its evaluation harness via mechanism-level skills and evolving tools, improving pass@1 over GRPO by up to 12.6 points on SWE-bench variants with a 9B base model.

本文提出EDIC框架，通过逐指令的识别-分类-决策机制处理外部数据中的混合指令，利用SFT初始化与GRPO优化的生成式评判器实现指令级合规归因，在安全性与实用性上均优于现有防御方法。

这篇论文提出 Concurrent Constraint Reinforcement (CCR)，将多个可独立验证的问题打包进单个受上限的补全中，使 RLVR 在 8,000 token 短窗口下仍保留正确性信号，相比 32k 窗口的 GRPO 每步训练 token 成本降低约 80%。

AlphaBet is an agentic reinforcement learning framework that trains a single LLM policy to search, reason, and report probabilities across multiple pre-resolution dates, reducing Brier scores on Forecast-Dojo and FutureX while showing that GRPO's standard-deviation normalization can distort proper-scoring-rule targets.