Research map / Reinforcement learning / RL from feedback
Tree Search for LLM Agent Reinforcement Learning
Yuxiang Ji, Ziyu Ma, Yong Wang, Guanhua Chen, Xiangxiang Chu, Liaoni Wu
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 |
|---|---|---|---|
| replace chain-based rollouts with tree search (Tree-GRPO) | GRPO, Qwen2.5-1.5b · multi-hop QA avg EM | EM: 11.3 -> 19.1 | Improved |
| replace chain-based rollouts with tree search | GRPO, Qwen2.5-3b · multi-hop QA avg EM | EM: 31.8 -> 36.8 | Improved |
| replace chain-based rollouts with tree search | GRPO, Qwen2.5-14b · multi-hop QA avg EM | EM: 41.8 -> 45.3 | Improved |
| replace chain-based rollouts with tree search | GRPO, Qwen2.5-14b · single-hop QA avg EM | EM: 55.1 -> 55.7 | Improved |
| use tree-based rollout at budget 2 per prompt | GRPO, Qwen2.5-3b · multi-hop QA avg EM | EM: 14.9 -> 31.6 | Improved |
| add tree rollout on web-agent QA | GRPO, Qwen2.5-14b · GAIA avg F1 | F1: 16.4 -> 21.0 | Improved |
| use only intra-tree advantage at low budget (M=2,N=2,L=1) (ablation) | Tree-GRPO · multi-hop QA, Qwen2.5-3b | avg EM: collapse | Hurt |
| add intra-tree advantage to inter-tree advantage (ablation) | Tree-GRPO · multi-hop QA, Qwen2.5-3b, budget 4 | avg EM: 33.8 -> 36.8 | Improved |
| add inter-tree advantage to intra-tree advantage (ablation) | Tree-GRPO · multi-hop QA, Qwen2.5-3b, budget 8 | avg EM: 34.6 -> 36.4 | Improved |
| use token/sentence-level tree nodes instead of agent-step nodes (ablation) | Tree-GRPO · multi-hop QA, Qwen2.5-3b | avg EM: 36.8 -> 22.2 | Hurt |
| vary tree structure: fewer trees M with more branches N (ablation) | Tree-GRPO · multi-hop QA, Qwen2.5-3b, budget 16 | avg EM: 37.3 -> 35.3 | Hurt |
| increase expansion iterations L instead of N (ablation) | Tree-GRPO · multi-hop QA, Qwen2.5-3b, budget 2 | avg EM: 36.8 -> 36.3 | No clear effect |
Closest papers by meaning
- Group-in-Group Policy Optimization for LLM Agent Training (NeurIPS 2025 · similarity 0.91)
- Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs (ICLR 2026 · similarity 0.91)
- Group-Graph Policy Optimization for Long-Horizon Agentic Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- Agentic Reinforced Policy Optimization (ICLR 2026 · similarity 0.89)
- Policy Guided Tree Search for Enhanced LLM Reasoning (ICML 2025 · similarity 0.89)
- Stratified GRPO: Handling Structural Heterogeneity in Reinforcement Learning of LLM Search Agents (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.
