Research map / Reinforcement learning / RL from feedback
ARLArena: A Unified Framework for Stable Agentic Reinforcement Learning
Xiaoxuan Wang, Han Zhang, Haixin Wang, Yidan Shi, Ruoyan Li, Kaiqiao Han, Chenyi Tong, Haoran Deng, Alexander K Taylor, Renliang Sun, Yanqiao Zhu, Jason Cong and 2 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 sequence-level clipping (GSPO) instead of token-level GRPO clipping (ablation) | GRPO · Avg over tasks | score: 46.16 -> 52.28 | Improved |
| use tolerant clipping (CISPO/SAPO) instead of hard clipping (ablation) | GRPO · ALFWorld | success rate (%): 62.36 -> 54.42 (CISPO), 25.16 (SAPO) | Hurt |
| add sequence masking of negative-advantage low-ratio sequences (ablation) | CISPO · ALFWorld | success rate (%): 54.42 -> 78.88 | Improved |
| add sequence masking of negative-advantage low-ratio sequences (ablation) | SAPO · ALFWorld | success rate (%): 25.16 -> 76.92 | Improved |
| increase KL coefficient to 0.05 (ablation) | CISPO · ALFWorld | success rate (%): 54.42 -> 38.46 | Hurt |
| increase KL coefficient to 0.05 (ablation) | SAPO · ALFWorld | success rate (%): 25.16 -> 48.05 | Improved |
| enlarge off-policy mini-update batch to 1024 (ablation) | CISPO · ALFWorld | success rate (%): 54.42 -> 21.59 | Hurt |
| add fine-grained environment-level advantage (GiGPO) (ablation) | GRPO · ALFWorld | success rate (%): 62.36 -> 81.09 | Improved |
| add entropy-based advantage reshaping (EMPG) (ablation) | GRPO · ALFWorld / WebShop | success rate (%): 57.91 vs 62.36 on ALFWorld; 64.32 vs 57.71 on WebShop | Mixed |
| add dynamic filtering (ablation) | GRPO · ALFWorld | success rate (%): 62.36 -> 49.58 | Hurt |
| add dynamic filtering (ablation) | GIGPO · WebShop | success rate (%): 56.55 -> 76.82 | Improved |
| use seq-mean-token-mean loss aggregation (ablation) | GRPO · Avg over tasks | score: 46.16 -> 39.23 | Hurt |
Closest papers by meaning
- Agentic Reinforced Policy Optimization (ICLR 2026 · similarity 0.90)
- Dispelling the Mirage of Progress in Offline MARL through Standardised Baselines and Evaluation (NeurIPS 2024 · similarity 0.88)
- Learning Agentic Policy from Action Guidance (NeurIPS 2026 · similarity 0.88)
- FightLadder: A Benchmark for Competitive Multi-Agent Reinforcement Learning (ICML 2024 · similarity 0.88)
- Group-Graph Policy Optimization for Long-Horizon Agentic Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
- HyperMARL: Adaptive Hypernetworks for Multi-Agent RL (NeurIPS 2025 · similarity 0.87)
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.
