← All news·2026-08-22·3 min read

World Models Don’t Have a Crystal Ball — Here’s Why They Often Fail

MENTIS is a framework that incorporates 6 “mental variables” — beliefs, attention, goals, intentions, emotions, norms, and social relationships — and improves the accuracy of world models (Sora/Genie) from 63.3% F1 to 87.9%.

airesearchworld-modelsagents

World models predict future states (e.g., Sora/Genie). But these world models don’t have human-like mental states such as beliefs, attention, goals, intentions, emotions, norms, and social relationships, so their predictions are often wrong.

MENTIS framework provides 6 “mental variables” (beliefs, attention, goals, intentions, emotions, norms, and social relationships) to improve world model performance from 63.3% F1 score to 87.9%.

Amazingly, even though the model used in MENTIS is weaker than the model used without MENTIS (84.9% vs 83.6%), MENTIS still performs better! Particularly in social interactions (+26.4 points). Human performance is nearly perfect (98.5%).

The majority of the rest of the mistakes are due to state transition simulations, which involve both physical and mental state changes. (80%)

In this paper, we demonstrate that unless your agent or robot understands what people want, it will still make mistakes, even if it knows everything about the physical environment around us. Let’s start adding “psychology” to world models for real-world applications!

Source: the-decoder.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids