Token Predictors Are Not Planners: Building Physically Grounded Causal Reasoners

Researchers introduce Causal-Plan-Bench and Causal-Plan-1M to shift embodied AI systems from linguistic token prediction toward physically grounded causal reasoning. The work demonstrates that leading models like Gemini 3 Pro struggle with genuine physical planning, while their Causal Planner model achieves 36.3% relative performance gains through million-scale causal training data.

This research addresses a fundamental limitation in current embodied AI systems: their tendency to rely on statistical language patterns rather than genuine physical understanding. Leading models optimize for next-token prediction, a metric that rewards linguistic fluency without guaranteeing accurate physical reasoning. The authors demonstrate this gap empirically, showing that Gemini 3 Pro achieves only 38.18 on their diagnostic benchmark despite being a state-of-the-art model. This distinction matters because autonomous systems deployed in physical environments require causal understanding—knowing not just what comes next linguistically, but what physically happens next.

The research builds on growing recognition that vision-language models alone are insufficient for embodied AI. Current benchmarks inadvertently incentivize shallow pattern matching over causal modeling. By constructing Causal-Plan-Bench with multi-stage verification across four causal dimensions and Causal-Plan-1M with explicit reasoning traces from egocentric videos, the authors establish evaluation standards that reward physical grounding. Their findings reveal a scaling law: training data quality and quantity in causal reasoning drive measurable gains in physical planning accuracy.

For the AI development community, this work signals that frontier models require architectural and training changes beyond scale. The Causal Planner's 45.28 performance represents meaningful progress, yet remains far from robust physical autonomy. This research likely influences how future robotics and embodied AI systems are trained and evaluated, particularly among teams prioritizing real-world deployment over benchmark optimization. The emphasis on causal reasoning over linguistic prediction reflects broader maturation in AI safety and reliability concerns.

• →Current frontier models prioritize linguistic token prediction over physical reasoning, limiting reliable autonomous planning.
• →Causal-Plan-Bench introduces specialized evaluation addressing four causal dimensions to measure genuine physical grounding.
• →Causal-Plan-1M dataset of one million annotated reasoning traces enables 36.3% relative performance improvement in next-state prediction.
• →Causal Planner model based on Qwen3-VL-8B demonstrates stronger physical planning than larger models like Gemini 3 Pro.
• →Research reveals scaling laws for causal training data, establishing clear performance gains as training corpus grows.