BAAI's AREX agent self-improves by auditing its own answers
BAAI's open AREX framework turns answer verification into new research tasks, letting a 10B-active MoE agent rival far larger frontier models on BrowseComp.
The Beijing Academy of Artificial Intelligence (BAAI) released AREX, a deep-research agent framework built around recursive self-improvement, and it topped Hugging Face's daily papers chart on Thursday with the day's highest upvote count.
Exploiting the discovery-verification gap
AREX's core insight is an asymmetry familiar from complexity theory: finding an answer that satisfies every constraint of a hard research question is expensive, but checking a proposed answer constraint-by-constraint is much cheaper. The framework runs two nested loops. An inner research loop gathers evidence and assembles provisional answers with confidence scores; an outer self-improvement loop then audits each answer against individual constraints, flags unresolved claims, and converts the gaps into targeted follow-up research problems for the next round. A learned update_context tool compacts the agent's interaction history into an "improvement state" — verified evidence, constraint status, open gaps, next steps — instead of dragging full transcripts through long sessions.
Small models, frontier results
BAAI instantiated the framework as AREX-Base, a mixture-of-experts model with 122 billion total and 10 billion active parameters, and AREX-Turbo, a 4-billion-parameter dense model. Across BrowseComp, WideSearch, DeepSearchQA, Humanity's Last Exam, GAIA, and xbench-DeepSearch-2510, the paper reports AREX-Base substantially beats comparable-scale baselines and stays competitive with much larger frontier systems, headlined by 82.5 percent on BrowseComp — a benchmark where early frontier agents struggled to crack 50.
Why it matters
Deep research has become the proving ground for agent architectures, and AREX makes a pointed argument: the returns now come from the harness, not the parameter count. If a 10B-active open model can audit and re-target its own work well enough to hang with frontier systems, the moat around proprietary research agents narrows — and the 'self-improving agent' pattern, which Tencent's Hyra-1.0 also pushed this week, starts to look less like a research curiosity and more like the emerging default for how serious agent systems get built.