Skip to content
conftrace_
2026 EMNLP EMNLP 2026

World-Model Evolving Agents

AbstractAgents learn from experience, distilling it into memories, skills, or abstractions that improve which action they propose. The consequence of that action is discovered only after execution, when a long-horizon error has already become irreversible. We introduce WEVA (World-model Evolving Agents), a framework that makes gained experience predictive. WEVA learns from interaction traces what information is predictive and how actions transform the resulting state, using a quasi-symbolic representation over entities, preconditions, and outcomes. Specifically, before acting, EVA predicts the states induced by candidate action sequences, discards those that violate learned preconditions, and executes the first action of the highest-scoring surviving branch. The agreement between predicted and realised states delivers a model-trust signal that regulates when lookahead is reliable and when further interaction is more informative. WEVA improves task success while executing fewer actions, reduces trajectories that terminate in externally verified constraint violations, and detects changes in environmental dynamics before they are reflected in task success. These results position WEVA as a step towards agents that predict the consequences of their decisions and verify them against what they hold about the world before acting.

More from EMNLP 2026

All EMNLP 2026 papers →