Tag
The paper introduces CivHarness, a replay-grounded system for generating long-horizon feedback in Freeciv, and CivTelescope, an LLM judge that predicts which game position leads to a better outcome, addressing 'scoreboard traps' where in-game scores mislead. Using CivTelescope as a dense reward (CivMarshmallow), agents achieve 8.94-point mid-game gains, 77% above the best baseline.
The paper proposes TARGRS, a test-time detoxification framework that learns token-level, transition-aware reward guidance over toxicity trajectories in LLM latent space, achieving up to 69.38% toxicity reduction on RealToxicityPrompts and extending to vision-language models with minimal inference overhead.
Rubricon is a pipeline from Handshake AI that automatically generates and refines discriminative rubrics for LLM evaluation and post-training without human labels or fine-tuning, achieving 9% better alignment with expert rubrics than prior synthetic methods and 4.5% better HealthBench performance when training Qwen3.5-9B.
CONCUR introduces a new advantage estimator for multi-task, multi-reward RL training of LLMs that projects gated auxiliary rewards onto scores concordant with primary rewards, outperforming baselines across math, code, instruction following, and writing tasks on three models.
本文提出 Discord 与 BRACE 两种方法，利用奖励分歧来校准 MCP 环境中工具使用语言模型的强化学习 rollout 预算分配，并在 Qwen3-VL 多个规模及 BFCL、NESTFUL 基准上验证其优于基线。
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%。
华为openJiuwen团队提出Reward Engine，通过构建由测试用例和评分细则组成的冻结RewardPack，为开放式工具智能体提供证据驱动的奖励信号，用于自修订与强化学习，在WorkBuddy Bench Office和Claw-Eval上显著提升通过一致性。
HarnessRL 提出一个与策略共同进化的训练框架，通过自适应任务采样、rollout 引导和 rubric 标准更新来解决开放式强化学习中的奖励失效问题，并发布了五领域评测集 Rubric-Mix-10k，在 4B 和 9B 模型规模上均优于最强基线。
CleanRM 使用少量干净反馈，通过最优传输校准逐样本噪声比例，并以噪声校正的代理损失训练奖励模型，在 RLHF 中实现无偏的干净风险估计并超越现有去噪基线。
Tencent Hy Frontier and academic collaborators introduce Terminal-RSI, a recursive self-improvement loop for terminal agents that alternates verified-trajectory rewriting with critic-inheriting policy refinement, lifting a 27B model's Terminal-Bench 2.1 success from 51.7% to 86.5% and reaching a new 19.8% SOTA on Terminal-Bench 4.0.
本文提出 MRTPO，通过基于回放的相关性度量将每个奖励的优势只分配给相关 token，从而改进多奖励强化学习训练，加速收敛并提升在 text-to-SQL、多约束指令遵循和工具调用任务上的性能。
ReMaRk 提出通过将生成列表与所有有效回复的标题联合评分来消除 GRPO 训练中的回复噪声，在保持或超越 Rank-GRPO 性能的同时显著降低 LLM 推荐模型的 RL 训练成本。
提出一种无需人工偏好标注的多模态奖励建模方法：在冻结的 Qwen3-VL 骨干网络的中间层上训练线性探针，通过扰动字幕对比进行训练，在多个基准上达到或超过 GPT-4o，且微调计算量仅为千分之一。
论文提出 Behavioral Self-Rewarding (BSR)，一种利用策略自身对同任务轨迹的调试行为进行两两比较并生成辅助奖励的奖励精炼方法，超越二元可验证奖励，在 SWE-bench Verified 和 SWE-bench-Live 上以 Qwen3-14B/32B 取得一致提升。
该论文提出 Tracking-Value PPO（TPPO），通过在对数几率空间对前缀值进行贝叶斯分解，构建累积值参数化以更好地跟踪语言模型推理中的值演化。TPPO 在 Qwen3-4B-Base 和 Qwen3-8B-Base 上分别比 PPO 提升 7.0 和 8.4 个数学推理分数点，并改善了分布外泛化能力。
The paper introduces LatentGRM, a latent evaluation framework that performs rubric-guided reward modeling in continuous latent space via semantics-preserving chunk compression, achieving competitive preference accuracy against SFT judges while compressing evaluation trajectories 8.9–9.2× and reducing inference time 6.1–7.0× at 4B/8B scales.
LatentGRM 提出了一种基于语义分块与压缩重建的潜在评估框架，通过在连续隐空间中完成 rubric 引导的成对偏好判断，将评估轨迹压缩 8.9–9.2 倍并把推理时间减少 6.1–7.0 倍，同时保持与显式 SFT 评委相当的偏好准确率。
HARPO is a reinforcement learning framework from Huawei Technologies that jointly optimizes hallucination mitigation and creative writing via a Hallucination-Aware Generative Reward Model, a selective activation mechanism, and a data curriculum, reducing hallucination rates while boosting creative-writing quality on Qwen models.
MatrixReward proposes constructing rewards for open-ended generation from a rollout-by-rubric win-rate matrix, using column spread and correlation statistics to derive data-dependent rubric weights and ideal profiles. Evaluated with Qwen3-8B on four benchmarks, it achieves 63.02 average score, outperforming the strongest baseline by ~2%.
作者扩展了 Cognition 的 SWE-2 奖励函数，通过调节 λₑ（即帕累托前沿在该努力水平处的导数）来主动引导帕累托前沿，而不仅仅是优化它，从而针对期望的分数/成本改进平衡进行目标化。