Meta's TreeFlex lets agents navigate their own search trees: 4B model hits 48.8% vs 41.5% for ReAct

dair_ai · x · 2026-10-15

Meta AI's new paper introduces TreeFlex, a framework where the agent itself decides where to search:

Problem: ReAct-style agents pile failed attempts and off-path observations into a single context, burying useful evidence.
Approach: The agent saves each intermediate state as a node in a search tree with short summaries. The same model acts in a Reflection role (picking which node to resume, branch, verify, or stop) and an Execution role (carrying out instructions). Trained jointly with a tree-aware GRPO variant using outcome and process rewards.
Results: At the same context limit, a 4B model reaches 48.8% on a deep-research benchmark vs 41.5% for flat ReAct and 42.8% for FoldAgent, with Pass@8 10 points higher. From the same 9B base with RL, TreeFlex resolves 54.9% of SWE-bench Verified vs 43.3% for FoldAgent.

More from coding & agent