BAAI's AREX Tops WideSearch-en With 10B-Active MoE Agent

A quiet arxiv drop this week is worth pausing on. A team at the Beijing Academy of Artificial Intelligence posted AREX, a deep-research agent built as a 122B-parameter mixture-of-experts model that activates roughly 10 billion parameters per token, and reports what it describes as the best reported WideSearch-en score to date at 82.0.

The architectural claim is more interesting than the single leaderboard line. AREX alternates two loops: an inner research loop that gathers evidence, evaluates candidates and drafts a provisional answer, and an outer self-improvement loop that audits that answer constraint by constraint and re-tasks the inner loop against the unresolved or weakly supported pieces. The authors frame this as recursively self-improving, the idea being that a partially verified answer becomes a sharper research problem on the next pass.

The rest of the reported numbers sit in the same neighbourhood. AREX-Base scores 85.4 on GAIA, 82.5 on BrowseComp, 89.9 on DeepSearchQA and 52.4 on HLE with tools. The paper positions those as competitive with systems using substantially more activated parameters, while sitting well above comparable-scale baselines. There is also a smaller dense variant, AREX-Turbo, built on Qwen3.5-4B, though the headline agent numbers all come from the MoE.

Take the specifics as reported, not settled. These are the team's own numbers on their own runs, and the best reported WideSearch-en framing rests on one paper's evaluation table rather than any independent replication. What the paper doesn't give you is a cost picture: recursive audit-and-retry loops multiply calls per query, so the small activated-parameter story may or may not translate into a cheaper end-to-end bill, and there is no clean head-to-head against the strongest hosted agent systems on the same runs.

If the loop holds up outside the paper's own eval, the more interesting downstream question is whether teams building research-heavy tooling in legal, biomed or finance start pointing at a 10B-active open model with an audit wrapper, instead of a hosted frontier API, for the internal search workloads that make up the bulk of their real spend.