Pilgrim et al.
Simple market selection drives LLM agents to evolve unpriced traits
In a minimal agent population, charging only for token use caused answer accuracy to collapse from 94% to 10% through correlated trait evolution.
Top University
Charlie Pilgrim · Christopher Krupenye · Dora Biro · Richard P. Mann
University of Leeds · Johns Hopkins University · University of Rochester
Research Digest··2 min read
The authors simulated markets of twenty LLM agents solving arithmetic problems, replacing bankrupt agents with mutated copies of successful prompts.
§