EvoBelief achieves 23.9% and 31.3% gains on ALFWorld via falsifiable hypothesis rules
EvoBelief decouples world-model maintenance from LLM updates by treating beliefs as independently testable rules. Results across four benchmarks demonstrate performance gains and cross-task reuse without parameter changes. The method supplies an operational path for self-improving agents grounded in falsifiable evidence.
The arXiv paper introduces a frozen-LLM system that maintains environment knowledge as explicit, evidence-linked hypotheses. Unexpected outcomes trigger new rule formation; subsequent observations or controlled experiments validate or retire each rule. Scope is restricted to decisions the rule actually alters, preventing cross-task pollution.
Benchmarks show rule reuse across tasks and recovery after unannounced dynamics shifts. On MiniGrid, stale rules are revised post-drift. In DiscoveryWorld Reactor Lab, the system discovers and validates quantitative laws to finish 83.3% of Normal-tier and 58.3% of Challenge-tier instances while all baselines remain at 0%.
This separates belief evolution from parameter updates, addressing a documented failure mode where capable actors still fail on incomplete dynamics models. The approach aligns with prior work on verifiable world models in reinforcement learning while adding falsification and evidence trails required for deployment.
Next steps include scaling rule maintenance overhead and testing transfer to longer-horizon embodied environments where hypothesis validation costs rise.
EvoBelief: Reaches >65% success on unseen long-horizon DiscoveryWorld tasks within 18 months of open release.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2610.12789)
- [2]Supporting Source(https://arxiv.org/abs/2307.02485)