What's this about?
AREX is a deep-research agent built on one asymmetry: finding an answer that satisfies many constraints at once is expensive, but checking a candidate answer decomposes into cheap, independent verifications. The agent uses its own verified results to improve its search strategy — recursively. It topped Hugging Face's paper leaderboard by a wide margin this week (140 upvotes).
What exactly is "discovery–verification asymmetry"?
Some problems are hard to solve but easy to grade — like a jigsaw puzzle: assembling it takes hours, confirming it's assembled takes a glance. If verification is cheap and trustworthy, an agent can generate many candidate answers, keep only the verified ones, and train itself on its own successes without human labels.
Why does it matter?
Self-improvement loops grounded in verifiable checks are the current bet for pushing agent capability without proportional human supervision — the same idea behind RLVR, and (per this weekend's debate) the very training philosophy critics blame for models that ace checks while drifting from human preference.