AREX Open app
AREX

Recursively Self-Improving Deep Research Agents.

Verification preserves established evidence and turns unresolved constraints into the next research objective.

AREX benchmark results across BrowseComp, GAIA, xbench-2510, DeepSearchQA, WideSearch, and HLE.
Six search, reasoning, and tool-use evaluations. AREX-Turbo · 4B dense   /   AREX-Base · 122B-A10B MoE
See the framework

Framework

Research. Verify. Improve.

AREX alternates an inner research loop with an outer self-improvement loop. Each provisional answer is audited constraint by constraint, then accepted, refined, or restarted.

AREX framework with an inner research loop and an outer self-improvement loop that accepts, refines, or restarts research based on verification.
Verified findings survive across rounds; unresolved constraints guide what happens next.
Method & models

Method in brief

A better research state, every round.

AREX uses verification as a transition—not a final filter. It preserves supported evidence, compresses the working context, and turns remaining uncertainty into a targeted next research problem.

  1. 01

    Research

    Search, read, and construct a provisional answer.

  2. 02

    Verify

    Audit each required claim against its evidence.

  3. 03

    Update

    Keep verified findings, open constraints, and the next plan.

  4. 04

    Recurse

    Accept, refine, or restart from the improved state.