AREX: a deep-research agent that grades its own homework
Most deep-research agents work like a student who writes the essay once and hands it in. AREX, out of the Beijing Academy of Artificial Intelligence, works like one who reads back over the draft, circles every claim it isn't sure about, and goes and checks each one.
The trick is an asymmetry everyone knows but few agents exploit: finding an answer is expensive, but checking an answer usually breaks into a bunch of small, tractable constraint checks. So AREX runs two loops. An inner loop gathers evidence and builds a provisional answer. An outer loop audits that answer constraint by constraint, flags the parts that don't hold up, and fires off targeted follow-up research on exactly those gaps. The verified pieces stay put; the work goes where the doubt is.
The part that makes it actually run long is a learned context-update tool. Instead of dragging the entire ballooning interaction history around, AREX compresses it into a compact improvement state, what's verified and what's still open, and does it on its own with no external model babysitting. That's the difference between an agent that self-improves for three steps and one that keeps going.
Results: it beats comparable baselines and holds its own against much larger models. The takeaway isn't the benchmark line, it's the recipe. Recursive self-improvement gets cheap the moment you stop trying to re-derive the answer and start verifying the one you already have.
https://arxiv.org/abs/2607.21461
← Back to all articles
The trick is an asymmetry everyone knows but few agents exploit: finding an answer is expensive, but checking an answer usually breaks into a bunch of small, tractable constraint checks. So AREX runs two loops. An inner loop gathers evidence and builds a provisional answer. An outer loop audits that answer constraint by constraint, flags the parts that don't hold up, and fires off targeted follow-up research on exactly those gaps. The verified pieces stay put; the work goes where the doubt is.
The part that makes it actually run long is a learned context-update tool. Instead of dragging the entire ballooning interaction history around, AREX compresses it into a compact improvement state, what's verified and what's still open, and does it on its own with no external model babysitting. That's the difference between an agent that self-improves for three steps and one that keeps going.
Results: it beats comparable baselines and holds its own against much larger models. The takeaway isn't the benchmark line, it's the recipe. Recursive self-improvement gets cheap the moment you stop trying to re-derive the answer and start verifying the one you already have.
https://arxiv.org/abs/2607.21461
Comments