EMNLP 2026 Paper: Training World Models for Behavior Consistency Cuts False Positives from 42.5% to 9.5%

机器之心 · wechat · 2026-09-08

A joint EMNLP 2026 paper from DUT, MBZUAI, PKU and Microsoft reveals an "inverted metric" problem in text-based world models: training to maximize textual similarity can actively push hallucinations toward decision-fatal errors. In controlled tests, F1, BERTScore, ROUGE-L and even GPT-4o-as-judge all rank outputs wrongly — a missing target product scores higher than missing two irrelevant items.

Method — BehR (Behavior Consistency Reward): a frozen reference agent scores the same action's log-probability under both real and predicted states; the gap serves as the reward, optimized with GRPO. No human preference labels needed.

Results across 16 configs (WebShop + TextWorld, two base models, four eval agents incl. GPT-4o/GPT-5): improvements in 13, no regressions. WebShop paired consistency for Qwen3-8B rose from 0.345 to 0.483; GPT-4o from 0.760 to 0.840. Offline evaluation false positives on TextWorld dropped from 42.5% to 9.5%. For lookahead planning, BehR lifted Qwen3-8B's WebShop success from 14.5% to 25.5%.

Key takeaway: a world model's value lies in producing consistent agent decisions, not textually faithful predictions. Code and models are open-sourced.

Original post →

More from coding & agent

coding & agent channel →