Skip to main content
Back

Causal Planner Beats Gemini 3 Pro at Physical Reasoning With 1M Training Traces

A new benchmark and 1M-sample training corpus expose how frontier VLMs fail at physical planning, and a fine-tuned 8B model beats Gemini 3 Pro by learning causal reasoning from egocentric video

  • The core finding: Leading VLMs score as low as 38.18 on physical planning tasks because they predict plausible language, not physical next-states.
  • New benchmark: Causal-Plan-Bench tests four causal dimensions (preconditions, effects, constraints, sequencing) designed to block language-prior shortcuts.
  • New dataset: Causal-Plan-1M provides 1 million causal reasoning traces from egocentric videos, annotated via a four-stage pipeline.
  • Causal Scaling Law: Scaling causal training data to 1M examples yields a 36.3% relative gain (33.22 to 45.28), beating Gemini 3 Pro.
  • Model: Causal Planner, fine-tuned on Qwen3-VL-8B, shows strong cross-benchmark generalization beyond its training distribution.
  • Key insight: The bottleneck for physical AI is not model size but causally-structured training data; the annotation pipeline is the reusable contribution.

Ask a frontier vision-language model to plan a sequence of physical actions -- say, sorting laundry before washing it -- and it will often produce something that sounds plausible but is physically wrong. It is not reasoning about cause and effect; it is pattern-matching on language. A new paper, "Token Predictors Are Not Planners", puts a number on exactly how bad this problem is, and then shows a concrete path to fixing it.

The core argument is simple but damning: current benchmarks for embodied planning reward models that are good at predicting the next likely word, not the next physical state. A model trained on internet text has seen millions of cooking recipes, repair manuals, and how-to guides. It has learned the statistical shape of action sequences. But knowing that "stir" typically follows "heat" is not the same as understanding that the pan must be hot before stirring changes anything. That distinction -- between linguistic plausibility and physical causality -- is what this work is designed to measure and train.

The gap nobody was measuring

Embodied AI -- the field of building agents that can perceive, plan, and act in physical environments -- has been making rapid progress on standard benchmarks. Models like Qwen3-VL score 63% on perception tasks, but that drops to 42% on planning, a consistent decline across architectures that highlights how much harder physical reasoning is than visual recognition. The trouble is that many existing benchmarks still let models cheat by leaning on language priors. If the answer to "what should you do next?" is always a word that commonly follows the previous action in text, a language model can score well without ever modeling the physical world.

The researchers argue this is a fundamental mismatch: