EvoBelief: Falsifiable World Models for Self-Improving LLM Agents

What this work does

EvoBelief maintains an external world model of falsifiable rules while keeping the underlying LLM frozen. Learned claims gain permission to affect decisions only after validation and can lose that permission when observations contradict them.