Research map / Reinforcement learning / Visual and representation RL
Scaling Agent Learning via Experience Synthesis
Zhaorun Chen, Zhuokai Zhao, Kai Zhang, Bo Liu, Qi Qi, Yifan Wu, Tarun Kalluri, Xuefei Cao, Yuanhao Xiong, Haibo Tong, Huaxiu Yao, Hengduo Li and 6 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 |
|---|---|---|---|
| add experience replay (ablation) | DreamGym · WebShop | success rate (%): 59.2 -> 63.9 | Improved |
| add experience replay (ablation) | DreamGym · WebArena | success rate (%): 9.7 -> 13.3 | Improved |
| add experience reasoning (ablation) | DreamGym · WebShop | success rate (%): 55.8 -> 63.9 | Improved |
| add experience reasoning (ablation) | DreamGym · WebArena | success rate (%): 7.3 -> 13.3 | Improved |
| add curriculum task generation (ablation) | DreamGym · WebShop | success rate (%): 57.3 -> 63.9 | Improved |
| add curriculum task generation (ablation) | DreamGym · WebArena | success rate (%): 7.3 -> 13.3 | Improved |
| train on synthetic experience only instead of 80K real transitions (GRPO) | Qwen2.5-7B GRPO · WebArena | success rate (%): 6.1 -> 12.7 | Improved |
| train on synthetic experience only instead of 80K real transitions (GRPO) | Qwen2.5-7B GRPO · ALFWorld | success rate (%): 79.8 -> 71.0 | Hurt |
| add sim-to-real RL with 5K real transitions after synthetic training (GRPO) | Llama-3.2-3B GRPO · WebShop | success rate (%): 59.3 -> 70.5 | Improved |
| add sim-to-real RL with 5K real transitions after synthetic training (PPO) | Llama-3.1-8B PPO · WebShop | success rate (%): 58.1 -> 63.9 | Improved |
| transfer DreamGym policy from WebShop/WebArena to ALFWorld | DreamGym policy · ALFWorld | success rate | Hurt |
Closest papers by meaning
- WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks (CVPR 2026 · similarity 0.89)
- Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.89)
- ML-Agent: Reinforcing LLM Agents for Autonomous Machine Learning Engineering (ICML 2026 · similarity 0.89)
- Learning Agentic Policy from Action Guidance (NeurIPS 2026 · similarity 0.88)
- Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments (ICLR 2026 · similarity 0.88)
- Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning (ICML 2026 · similarity 0.88)
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.
