A new AI system just helped researchers cross five open problems off the list, by arguing with itself.
Researchers built Cogentic, a multi-agent harness that runs an iterative prove-verify loop instead of asking one model for one answer. An orchestrator splits a pool of independent provers across competing proof directions, built on top of Google's Gemini model, while several specialized verifier components adversarially check each candidate step. Confirmed intermediate results get locked into what the team calls a persistent verified ledger, so later rounds build on solid ground rather than re-deriving earlier work. Using this setup, Cogentic produced novel results on five open problems spanning online learning, auction theory, and mechanism design, with each result independently checked by outside domain experts and written up in companion papers.
The interesting part isn't the five results themselves, it's the shift in method. Rather than a single clever guess from a chatbot, this is sustained, adversarial exploration of a problem space over many rounds, closer to how a research group grinds through a hard proof than how a model answers a prompt. That's a different bet than prior formal-math systems, which mostly target competition-style problems with clean, checkable answers rather than genuinely open research questions.
Five results across three adjacent corners of theoretical economics is a real, modest haul, not a sweeping one, and it still took human experts to sign off on every single result.