Research map / Reinforcement learning / Visual and representation RL
RLVMR: Reinforcement Learning with Verifiable Meta-Reasoning Rewards for Robust Long-Horizon Agents
Zijing Zhang, Ziyang Chen, Mingxiao Li, Zhaopeng Tu, Xiaolong Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add verifiable meta-reasoning rewards to GRPO-style RL | Qwen2.5-1.5B on ALFWorld · ALFWorld L2 | success rate (%): 29.7 -> 56.3 | Improved |
| use RLVMR instead of GiGPO | Qwen2.5-7B on ALFWorld · ALFWorld L2 | success rate (%): 67.2 -> 83.6 | Improved |
| use RLVMR instead of GiGPO on ScienceWorld seen split | Qwen2.5-7B · ScienceWorld L0 | success rate (%): 53.4 -> 67.2 | Improved |
| remove outcome-based trajectory advantage (ablation) | RLVMR 1.5B · ALFWorld L2 | success rate (%): 12.5 -> 56.3 | Improved |
| add meta-reasoning advantage signal (ablation) | RLVMR 1.5B · ALFWorld L2 | success rate (%): 45.3 -> 56.3 | Improved |
| add cold-start SFT on 200 trajectories (ablation) | RLVMR 1.5B · ALFWorld L2 | success rate (%): 40.6 -> 56.3 | Improved |
| use RLVMR instead of GRPO to reduce repetitive actions | Qwen2.5-7B · ALFWorld L0 | repetitive action rate (%): 21.5 -> 2.3 | Improved |
| use RLVMR instead of GRPO for episode length | Qwen2.5-7B · ALFWorld L2 | avg actions: 21.7 -> 15.4 | Improved |
| scale model from 1.5B to 7B under GRPO | GRPO · ALFWorld L2 | repetitive action rate / success: success 29.7 -> 52.3; repetitive 27.1 -> 31.2 | Mixed |
| SFT on expert trajectories vs GRPO on unseen L2 | Qwen2.5-7B · ALFWorld L2 | success rate (%): 52.3 (GRPO) -> 37.5 (SFT) | Hurt |
Closest papers by meaning
- The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models (NeurIPS 2026 · similarity 0.90)
- Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning (ICLR 2026 · similarity 0.90)
- SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search (ICLR 2026 · similarity 0.89)
- SAGE: Shaping Anchors for Guided Exploration in RLVR of LLMs (ICML 2026 · similarity 0.89)
- Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs (ICLR 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
