Tag
InteractiveWorld introduces the first verifiable simulation environment and InteractiveBench benchmark for evaluating whether MLLMs can reliably guide humans through physical tasks, finding that even frontier models like Claude Fable 5.1 achieve at most 20-25% on long-horizon guidance and planning tasks.
InteractiveWorld 提出了首个用于人机闭环交互的可验证模拟环境，并配套发布 InteractiveBench 基准（含六大类 120 个多轮任务），用于评估 MLLM 引导人类完成物理任务的能力；实验显示 12 个前沿 MLLM 仍远未可靠，最强模型在长程任务上得分仅约 20-25%。
This paper explains why majority-vote self-rewards match verifiable-reward RL by showing that, at the batch level, incoherent label errors cancel out so the update direction aligns with the gold gradient; it finds error coherence—not error rate—determines when self-generated labels can replace verification.
提出 Replay-Selected Hindsight Distillation (RSHD) 方法，让模型从完全失败的推理组中恢复监督信号：策略对失败尝试撰写纠错笔记并重新求解，仅当重放探索成功时才将该笔记蒸馏回普通提示策略中，在多个模型族和数学/代码训练域上一致优于 GRPO。
论文提出 Gibbs-Gap Policy Optimization（GGPO），一种针对可验证奖励的离线/异步 RL 方法，通过最小化 KL 调整奖励的 log-mean-exp 对比来避免高方差重要性权重，并在数学推理基准上优于 GRPO 和 AGRO。
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。
评论介绍了 UCLA 与 Adobe Research 的 Tev-Gym 工作：通过构建可验证的长任务环境与稠密中间进度奖励来训练长时程 Agent，Tev-27B 在 ALE-CM 计算与数学子集上以 Pass@5 比肩 GPT-6 Astra 与 Claude Opus 5，Tev-320B 则将 GLM-5.3-Flash 在 Terminal-Bench 4.0-CM 上的 Pass@1 从 29% 提升至 37%。
论文研究了LLM在可验证奖励强化学习（RLVR）下的优化几何特性，发现梯度能量会逐渐集中到不断收缩的有效子空间中，进而提出自适应秩优化器Ares，根据梯度谱能量分布动态调整更新秩，在多个模型规模和RLVR任务上取得优于主流优化器的表现。
KaliBench 介绍了一个针对 Kali Linux 的细粒度基准与数据集，涵盖 8,504 个自然语言到 CLI 命令的配对，用于评估 LLM 生成可执行网络安全工具命令的能力，并通过无运行时的可验证奖励支持模型训练。
The paper identifies an off-policy asymmetry in on-policy distillation, where the teacher must supervise student-generated prefixes it was not trained on, and proposes SCOUT, a co-training framework that adapts the teacher via RL with verifiable rewards, consistently improving distillation across configurations, scales, and reasoning domains.
Introduces Hindsight-Divergence Localization (HDL), a method for efficient reinforcement learning with verifiable rewards that reduces generation costs and improves performance across math, code, and agent tasks by selecting branch points based on token log-likelihood changes.
This paper tests Reinforcement Learning with Verifiable Rewards (RLVR) on small language models using retrieval-augmented generation, achieving a 3.8-fold gain without distillation and emphasizing the need for reward design tailored to small models.
EAGER is a reinforcement learning framework that improves generative event extraction through fine-grained verifiable rewards and schema-contrastive advantage estimation, outperforming prompting, fine-tuning, and prior RL methods on seven benchmark datasets.
VERPO introduces a framework for verified evidence regularized policy optimization in language models, improving performance on scientific reasoning and tool-use tasks by treating evidence as a proposal for policy correction while maintaining outcome objectives.
VBVR-Pro introduces a closed-loop testbed for scalable and verifiable native visual reasoning through generation, featuring task scaling, verifiable rewards, and mechanism studies across diverse visual substrates.
This talk by Will Brown of Primordial AI discusses techniques for scaling Reinforcement Learning to complex, real-world tasks where rewards are not verifiable, using methods like anchoring, LLM judges, and simulation.
This paper introduces SARA, a sequential adaptive rollout allocation method for RLVR that abandons saturated groups early and reallocates the budget, achieving comparable accuracy with 22% fewer rollouts than dynamic sampling and up to 67% savings when combined.
This paper proposes RLSVR, a task-transformation paradigm that extends reinforcement learning with verifiable rewards to open-ended LLM tasks by creating self-verifiable proxy environments, instantiated via the SpyRL multi-agent self-play framework, showing gains on summarization, creative writing, and math reasoning.
This paper diagnoses 'pass@k inversion' in reinforcement learning with verifiable rewards (RLVR), where training improves one-shot accuracy but reduces performance under repeated sampling, particularly on boundary prompts where correct trajectories are rare. It proposes a per-problem base anchoring (PBA) method to mitigate this issue.
Introduces Surrogate Latent Policy Optimization (SLPO) to apply outcome-reward RL to autoregressive latent reasoners, enabling test-time scaling and variable-horizon policies that improve accuracy on harder instances.