EvoBelief preprint reports agent gains through revisable environment rules

An arXiv preprint reports that EvoBelief improves AI agents' task performance by updating their knowledge of how an environment works while keeping the language model frozen. On ALFWorld without actor-side transition memory, success improved over a matched Frozen World Model by 23.9% with Claude Haiku 4.5 and 31.3% with GPT-5.4. EvoBelief forms testable hypotheses from unexpected outcomes and relevant unknowns. Observations and controlled experiments validate rules before they guide decisions; counterexamples narrow or retire them. In DiscoveryWorld's Reactor Lab, it completed 83.3% of Normal-tier and 58.3% of Challenge-tier instances, versus 0% for each evaluated baseline. The reported gains come from benchmark environments.