LLM agents fail to converge in double auctions, allocate less efficiently than humans
WillRinehart · x · 2026-09-22
An arXiv paper replicates classic double-auction experiments with LLM agents replacing humans. Markets populated by LLMs converge slower or not at all toward equilibrium, yielding less efficient allocations, with substantial heterogeneity across model families and market roles. Lexical analysis of CoT traces shows trade execution correlates with a shift from strategic reasoning to urgency. The testing framework is publicly released.
More from AGI Musings
- Gebru and Bender: this summer's AI breakthroughs are marketing, not magic — nordicinst · 2026-09-22
- Life advice for the AI age: keep judgment, protect attention, don't measure worth by irreplaceability — CurieuxExplorer · 2026-09-22
- Ex-Pentagon AI Official: Adopting AI Is an Institutional Problem, Not a Model Problem — ChinaTalk · 2026-09-22
- Box CEO Aaron Levie bets big on personal agents, as Meta agent spending could push ARPU past $227 — inductionheads · 2026-09-22
- The week labs asked to slow down: Claude does 26% of Anthropic's work, Gemini breached 3 firms — alex_verem · 2026-09-22
- White-collar work has ~3 years left, argues ChrisGPT, before agents abstract it away — ChrisGPT · 2026-09-22