BAAI’s AREX agent uses recursive verification to improve deep research answers
BAAI · hf · 2026-07-24
Key idea
AREX is a recursively self-improving deep research agent that separates discovery from verification: it first gathers evidence and drafts a provisional answer, then runs an outer loop that audits the answer constraint by constraint and launches targeted follow-up research on unresolved claims.
What is new
- Introduces an autonomous context-update tool that compresses long interaction history into a compact improvement state, preserving verified evidence and open constraints without relying on another model.
- Trains the agent with verified synthetic tasks and high-quality trajectories using agentic mid-training and long-horizon reinforcement learning.
- Uses step emphasis to address sparse rewards, focusing learning on moments when decisive evidence is found or a bad research path is corrected.
Results
- Instantiates both a dense 4B model and a 122B-A10B MoE model.
- Reports strong results on BrowseComp, WideSearch, DeepSearchQA, Humanity's Last Exam (HLE) and other reasoning/tool-use benchmarks.
- The paper says AREX beats comparable-scale baselines and remains competitive with models that activate far more parameters.
Related event: BAAI Introduces AREX: A Recursive Self-Improving Research Agent(2 posts)→
More from Research
- MaP-WAM tackles non-Markovian robot manipulation with memory-grounded planning — Sizhe Zhao · 2026-09-11
- Negative Self-Distillation improves LLM reasoning by avoiding flawed reasoning paths — Rongcan Pei · 2026-09-11
- ~25% of NBER economics papers now contain AI-generated text, study finds — JeremyNguyenPhD · 2026-09-11
- DeepMind-led paper makes design docs the source of truth, code disposable — SMART regenerates in 1.5-3h for ~$100 — Roger_M_Taylor · 2026-09-11
- GameWorld wins Best Paper Runner-Up at ECCV 2026 Multimodal Digital Agents Workshop — MikeShou1 · 2026-09-11
- Yann LeCun live at ECCV on World Models — Weak_Assistance_5261 · 2026-09-11