BehR: more realistic world models can make agents worse, EMNLP paper finds

jiqizhixin · x · 2026-09-17

A Dalian University of Technology paper accepted to EMNLP 2026 shows text-based world models optimizing for "state consistency" (how closely predicted text matches the real environment) can actually lead agents further astray. BehR shifts the objective to behavior consistency: whether the agent makes the same decisions in the predicted environment as in the real one. Across 16 experimental configurations, BehR improves trajectory-level consistency in 13, ties in 3, and never regresses.

Original post →

More from AGI Musings

AGI Musings channel →