AI Paper Atlas

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

ICLR 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants

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 changedWhereMeasuredEffect
replace chain-based rollouts with tree search (Tree-GRPO)GRPO, Qwen2.5-1.5b · multi-hop QA avg EMEM: 11.3 -> 19.1Improved
replace chain-based rollouts with tree searchGRPO, Qwen2.5-3b · multi-hop QA avg EMEM: 31.8 -> 36.8Improved
replace chain-based rollouts with tree searchGRPO, Qwen2.5-14b · multi-hop QA avg EMEM: 41.8 -> 45.3Improved
replace chain-based rollouts with tree searchGRPO, Qwen2.5-14b · single-hop QA avg EMEM: 55.1 -> 55.7Improved
use tree-based rollout at budget 2 per promptGRPO, Qwen2.5-3b · multi-hop QA avg EMEM: 14.9 -> 31.6Improved
add tree rollout on web-agent QAGRPO, Qwen2.5-14b · GAIA avg F1F1: 16.4 -> 21.0Improved
use only intra-tree advantage at low budget (M=2,N=2,L=1) (ablation)Tree-GRPO · multi-hop QA, Qwen2.5-3bavg EM: collapseHurt
add intra-tree advantage to inter-tree advantage (ablation)Tree-GRPO · multi-hop QA, Qwen2.5-3b, budget 4avg EM: 33.8 -> 36.8Improved
add inter-tree advantage to intra-tree advantage (ablation)Tree-GRPO · multi-hop QA, Qwen2.5-3b, budget 8avg EM: 34.6 -> 36.4Improved
use token/sentence-level tree nodes instead of agent-step nodes (ablation)Tree-GRPO · multi-hop QA, Qwen2.5-3bavg EM: 36.8 -> 22.2Hurt
vary tree structure: fewer trees M with more branches N (ablation)Tree-GRPO · multi-hop QA, Qwen2.5-3b, budget 16avg EM: 37.3 -> 35.3Hurt
increase expansion iterations L instead of N (ablation)Tree-GRPO · multi-hop QA, Qwen2.5-3b, budget 2avg EM: 36.8 -> 36.3No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy