ResearchUPDATE

AREX proposes a research agent that verifies answers and searches again

AREX audits a provisional answer constraint by constraint, then launches targeted follow-up research for claims that remain unresolved.

07/24/20261 sources reviewed
Quick summary
  • An inner loop gathers evidence and drafts an answer; an outer loop audits unresolved claims.
  • A learned context-update tool compresses long histories into verified evidence and remaining constraints.
  • The authors report that 4B and 122B-A10B MoE variants outperform comparable baselines on several search and reasoning benchmarks.
WHAT HAPPENED

What happened?

AREX audits a provisional answer constraint by constraint, then launches targeted follow-up research for claims that remain unresolved.

The architecture exploits the idea that verifying a candidate can be easier than discovering it. Independent evaluation is still needed on changing web content, conflicting sources, and restricted access beyond synthetic training and benchmarks.

WHY IT MATTERS

Why does it matter?

Instead of simply searching longer, the approach repeatedly verifies individual claims to improve the accuracy and efficiency of long-horizon research agents.

WHO SHOULD CARE

Who should care?

AI researchersSearch developersKnowledge workers
AIZIGOO VIEW

AIZIGOO view

This is a July 23, 2026 preprint; benchmark results are reported by the authors.