Self‑evolving LLM agents usually improve their actions by accumulating experience, reflecting, or acquiring new skills. Yet even procedurally capable agents can fail when their beliefs about environment dynamics are incomplete, incorrect, or outdated. This work explores a complementary self‑evolution axis: evolving those beliefs during deployment.\ \ EvoBelief represents environment knowledge as falsifiable hypotheses. Unexpected outcomes or decision‑relevant unknowns trigger hypothesis formation. Subsequent observations and controlled experiments provide independent validation; only after sufficient evidence does a rule guide predictions or plans within its scoped applicability. Counterexamples narrow or retire rules as the environment changes. The language model itself remains frozen, and each rule is linked to its evidence and the decisions it influences.\ \ Four benchmarks were used for evaluation:\
- ALFWorld (without actor‑side transition memory): compared to a matched Frozen World Model, success improves by 23.9% with Claude Haiku 4.5 and by 31.3% with GPT‑5.4.\
- MiniGrid: revises stale rules and recovers after unannounced dynamics changes.\
- DiscoveryWorld’s Reactor Lab: discovers and validates quantitative laws, achieving 83.3% completion on Normal‑tier and 58.3% on Challenge‑tier instances, whereas all evaluated baselines achieve 0%.\ \ These results demonstrate self‑improvement through testable, revisable environment knowledge rather than model‑parameter updates.\ \ Review