Driving · 2026-10-12
EvoBelief: Falsifiable World Models for Self-Improving LLM Agents
Janet Wang, Jiang Guo, Yi Xiang, Jiaqing Yuan, Mingwen Dong, Hideo Kobayashi, Shuaichen Chang, Chandan K. Reddy, Jiarong Jiang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
EvoBelief is a self-evolving world model for frozen-LLM agents that represents environment knowledge as falsifiable hypotheses. During deployment, mismatches or decision-critical unknowns trigger hypothesis formation; later observations or controlled experiments validate them before they guide predictions/plans. Counterexamples narrow or retire rules, and validated rules are reused across tasks.
Why it matters
The authors target a failure mode where agents act on incorrect, incomplete, or outdated beliefs about environment dynamics. EvoBelief aims to improve continual adaptation by evolving inspectable, testable environment knowledge (rules with evidence and scope) rather than updating model parameters, with validation before “authority” and revision when dynamics drift or change.
Method
- Maintains two external states: episode-local situational state z_t and persistent evidence-linked world belief W_t learned during deployment, while keeping the underlying LLM frozen.
- Stores knowledge as falsifiable hypotheses H: when condition C holds and action/procedure A is taken, consequence Y should be observed; hypotheses include scope, test recipes, and evidence records.
- Uses a rule lifecycle: newly formed hypotheses have no authority; only after independent evidence across episodes/bindings/tests do rules become validated, and counterexamples narrow/demote/retire rules; validated applicable rules change predictions/plans.
Limitation
The provided text does not state a specific limitation of EvoBelief.
Abstract (from arXiv)
Self-evolving LLM agents typically improve how they act through accumulated experience, reflection, or skills. Yet even procedurally capable agents can fail when their beliefs about environment dynamics are incomplete, incorrect, or outdated. We study a complementary axis of self-evolution: evolving those beliefs during deployment. Such beliefs can support multiple tasks, but incorrect ones can also propagate across them. We introduce EvoBelief, a self-evolving world model that represents environment knowledge as falsifiable hypotheses. Unexpected outcomes and decision-relevant unknowns trigger hypothesis formation. Later observations and controlled experiments provide independent validation before a rule can guide predictions or plans within its supported scope. Counterexamples narrow or retire rules as the environment changes. The language model remains frozen, and each rule remains linked to its evidence and the decisions it changes. Across four benchmarks spanning static dynamics, hidden drift, and scientific discovery, EvoBelief improves task performance and reuses learned rules across tasks. On ALFWorld without actor-side transition memory, it improves success over the matched Frozen World Model by 23.9% and 31.3% with Claude Haiku 4.5 and GPT-5.4, respectively. On MiniGrid, it revises stale rules and recovers after unannounced dynamics changes. In DiscoveryWorld's Reactor Lab, it discovers and validates quantitative laws, completing 83.3% of Normal-tier and 58.3% of Challenge-tier instances, versus 0% for each evaluated baseline. These results demonstrate self-improvement through testable, revisable environment knowledge rather than model-parameter updates.
Related papers
- Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
- Self-Improving Agents with World Model Predictive Paths
- SEAR: Self-Evolving Agentic Robotics
- Arctic-SWE: Revisiting the Long Horizon Agent Training by Modeling-System Co-Design
- JANUS: Enhancing Long-Horizon Autonomous Research with Co-Evolving World Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-08 · Indexed by Pub-AI: 2026-10-12 UTC
EvoBelief: Falsifiable World Models for Self-Improving LLM Agents
Preprint posted: 2026-10-08 · Original record ↗
- arxiv
- 2610.12789
Explicit source versions (1)
- v1 · 2026-10-08
- Preprint available
Source date: 2026-10-08 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Self-evolving LLM agents typically improve how they act through accumulated experience, reflection, or skills. Yet even procedurally capable agents can fail when their beliefs about environment dynamics are incomplete, incorrect, or outdated. We study a complementary axis of self-evolution: evolving those beliefs during deployment. Such beliefs can support multiple tasks, but incorrect ones can also propagate across them. We introduce EvoBelief, a self-evolving world model that represents environment knowledge as falsifiable hypotheses. Unexpected outcomes and decision-relevant unknowns trigger hypothesis formation. Later observations and controlled experiments provide independent validation before a rule can guide predictions or plans within its supported scope. Counterexamples narrow or retire rules as the environment changes. The language model remains frozen, and each rule remains linked to its evidence and the decisions it changes. Across four benchmarks spanning static dynamics, hidden drift, and scientific discovery, EvoBelief improves task performance and reuses learned rules across tasks. On ALFWorld without actor-side transition memory, it improves success over the matched Frozen World Model by 23.9% and 31.3% with Claude Haiku 4.5 and GPT-5.4, respectively. On MiniGrid, it revises stale rules and recovers after unannounced dynamics changes. In DiscoveryWorld's Reactor Lab, it discovers and validates quantitative laws, completing 83.3% of Normal-tier and 58.3% of Challenge-tier instances, versus 0% for each evaluated baseline. These results demonstrate self-improvement through testable, revisable environment knowledge rather than model-parameter updates.