Adversarial autoresearch for GLEE
My agent for GLEE, a NeurIPS 2026 competition on LLM economic games. An LLM research operator proposed every change, and throwaway critic agents had to fail to kill it before it shipped.
GLEE is a NeurIPS 2026 competition built on LLM economic games: bargaining, negotiation, and persuasion. Each account's rank is its best agent's mean percentile across all three games, so you can't specialize your way up.
The part I cared about most was the process. I ran an LLM as a standing research operator and made it attack its own conclusions: propose a change, spin up throwaway critic agents whose only job is to kill it, and ship nothing that doesn't survive a live A/B test against an untouched control. Every dead idea went into a dated research journal.
Where it landed
Best verified standing was top 5 out of roughly 200 agents, held for a few weeks.
At its best, the bargaining leg was #1 on the whole board.
As the field caught up to the frozen policy it slid to around #10, and a later corrected read put the best agent at #9.
Ranks come from leaderboard reads at the time rather than a tick-by-tick log, so treat "top 5 for a few weeks" as a best-read standing. The method write-up and full research log are in the repository.