Token Predictors Are Not Planners: Building Physically Grounded Causal Reasoners

2026-06-01

AI Summary

This work addresses the limitation of existing embodied vision-language planning models, which overly rely on linguistic statistical priors and lack explicit modeling of physical causal relationships, thereby hindering genuine physical autonomy. To overcome this, the study proposes a paradigm shift toward physics-driven causal reasoning and introduces the first framework that transitions from language token prediction to explicit causal planning. Key contributions include the development of Causal-Plan-Bench—a high-fidelity diagnostic benchmark—the release of Causal-Plan-1M, a million-scale corpus for causal reasoning, and the training of a causal planner based on Qwen3-VL-8B endowed with internalized physical logic. Experiments demonstrate that the proposed method substantially outperforms current state-of-the-art models, achieving a 36.3% relative performance gain (from 33.22 to 45.28) on in-domain and cross-benchmark evaluations, surpassing even Gemini 3 Pro (38.18).

Abstract

Current benchmarks for embodied vision-language planning often favor linguistic next-token prediction over physically grounded next-state reasoning. This rewards models that mimic statistical language priors rather than track causal dependencies, reducing physical planning to shallow sequence modeling. We argue that reliable physical autonomy requires a shift from linguistically grounded token prediction toward physically grounded causal reasoning. To this end, we introduce Causal-Plan-Bench, a high-fidelity diagnostic suite curated through multi-stage verification to evaluate embodied planning across four causal dimensions. We also construct Causal-Plan-1M, a million-scale corpus of explicit reasoning traces produced by a four-stage annotation pipeline over egocentric videos. Extensive evaluation shows that leading models still struggle to demonstrate genuine physical agency, with Gemini 3 Pro reaching only 38.18 on our benchmark. In contrast, our training recipe enables Causal Planner, built on Qwen3-VL-8B, to internalize physical logic for more accurate next-state estimation. The model achieves strong in-domain performance and cross-benchmark generalization, and reveals a Causal Scaling Law: scaling causal training data to one million instances yields a 36.3% relative gain, from 33.22 to 45.28. Overall, our work provides a concrete step toward turning agents from superficial token predictors into physically grounded causal reasoners.

Problem

• embodied planning
• causal reasoning
• physical grounding
• token prediction
• vision-language models

Innovation

• causal reasoning
• embodied planning
• physical grounding
• Causal-Plan-Bench
• Causal Scaling Law