Research map / Reinforcement learning / RL from feedback
RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments
Zhiyuan Zeng, Hamish Ivison, Yiping Wang, Lifan Yuan, Shuyue Stella Li, Zhuorui Ye, Siting Li, Jacqueline He, Runlong Zhou, Tong Chen, Chenyang Zhao, Yulia Tsvetkov and 5 more
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 |
|---|---|---|---|
| use adaptive difficulty instead of static difficulty ranges ([0,1], [0,20], [0,100]) (ablation) | Qwen2.5-7B-Base on Sorting and Multiplication · held-out OOD (50 environments) and ID accuracy | accuracy / effective prompt ratio | Improved |
| use static difficulty [0,1] with low upper bound (ablation) | Qwen2.5-7B-Base · Sorting / Multiplication | effective prompt ratio | Hurt |
| use adaptive difficulty instead of static [0,20] on 256 joint environments (ablation) | Qwen2.5-7B, R1-Distill-1.5B, DeepScaleR-1.5B, ProRL-1.5B-v2 · OOD accuracy on 50 held-out environments | accuracy | Improved |
| increase number of training environments (1, 4, 16, 256) (ablation) | RLVE · OOD accuracy on 50 held-out environments | accuracy | Improved |
| add more problems from a single environment (more data volume) (ablation) | RLVE · held-out environments | accuracy | No clear effect |
| continue RL with RLVE on 400 environments instead of original ProRL data | ProRL-1.5B-v2 · six reasoning benchmarks average | absolute improvement: 0.49% -> 3.37% | Improved |
| train on RLVE-GYM instead of DeepMath-103K | OpenThinker3-1.5B · six reasoning benchmarks average | average performance: about 2% absolute | Improved |
| train on RLVE-GYM instead of DeepMath-103K | OpenThinker3-1.5B · AIME 2024/2025 | accuracy: peak about one point above DeepMath-103K | No clear effect |
Closest papers by meaning
- Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · similarity 0.93)
- Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.92)
- Beyond Verifiable Rewards: Scaling Reinforcement Learning in Language Models to Unverifiable Data (NeurIPS 2025 · similarity 0.91)
- The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models (NeurIPS 2026 · similarity 0.91)
- Reinforcement Learning for Reasoning in Large Language Models with One Training Example (NeurIPS 2025 · similarity 0.91)
- RL Tango: Reinforcing Generator and Verifier Together for Language Reasoning (NeurIPS 2025 · similarity 0.91)
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.
