Meta's TreeFlex Turns Agent Context Into a Search Tree: 4B Model Hits 48.8% on BrowseComp-Plus
armand_ruiz · x · 2026-10-20
A new Meta AI paper introduces TreeFlex, which structures an agent's context as a search tree instead of one long context:
- Method: each intermediate state becomes a node with a short summary; a single LLM plays two roles — Reflection surveys the tree and picks a node to resume with an instruction, Execution works that path with task tools — freely deepening, branching, revisiting, verifying or stopping, with no fixed traversal order or per-node scoring.
- Training: a tree-aware GRPO variant jointly trains navigation, actions, checkpointing and proposal review, with outcome plus process rewards.
- Results: a 4B model reaches 48.8% on BrowseComp-Plus (vs 41.5% ReAct, 42.8% FoldAgent at the same active-context limit) and 54.9% on SWE-bench Verified vs FoldAgent's 43.3%.
The authors caution it's a small-scale, single-pass RL early signal. The takeaway: the harness is the product.
More from coding & agent
- Google open-sources Harness², turning any agent harness into a self-improving system — rez0__ · 2026-10-20
- Fo MCP gives an AI assistant emotions and tools to control real-world devices — khant_dev · 2026-10-20
- Fo MCP Launches: Give Your AI Agent Control of Lamps, Vacuums and CLIs — julianweisser · 2026-10-20
- Claude Code-built game got 664 plays and under 1% day-two retention — chongdashu · 2026-10-20
- "It's still OK to enjoy programming": dev pushes back on AI-era coding doom — facontidavide · 2026-10-20
- Gartner: 70% of Enterprises Will Drop Vendor-Built Agentic AI by 2028 Over 'FDE Washing' — alex_verem · 2026-10-20