Token Predictors Are Not Planners: Building Physically Grounded Causal Reasoners
Zheng Lu, Mingqi Gao, Qinlei Xie, Wanqi Zhong, Hanwen Cui, [0.46em] Heng Cao, Zirui Song, Yifan Yang, Chong Luo, Bei Liu, Yiming Li, [0.58em] Tsinghua University Microsoft Research Asia MBZUAI [0.25em] https://github.com/THUSI-Lab/Causal-Reasoner · 2026
Read Token Predictors Are Not Planners: Building Physically Grounded Causal Reasoners by Zheng Lu, Mingqi Gao, Qinlei Xie, Wanqi Zhong, Hanwen Cui, [0.46em] Heng Cao, Zirui Song, Yifan Yang, Chong Luo, Bei Liu, Yiming Li, [0.58em] Tsinghua University Microsoft Research Asia MBZUAI [0.25em] https://github.com/THUSI-Lab/Causal-Reasoner (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-01810.txt.
Introduction
Lu and co-authors argue that word guessing is not real planning. Robot and agent plans need a grasp of cause and effect. Tests today reward models that mimic text patterns. Such models can score well with no grasp of physics. The team seeks a test that shows this gap.
They build Causal-Plan-Bench, a suite with four cause types. Checks pass through many stages of review. They also build a large corpus of cause traces. The corpus holds one million records from first-person video. Their Causal Planner model trains on that corpus. The authors report a 36.3 percent gain over its base. Top rival models still score low on true agency.
Catalog record
| CREATOR | Zheng Lu, Mingqi Gao, Qinlei Xie, Wanqi Zhong, Hanwen Cui, [0.46em] Heng Cao, Zirui Song, Yifan Yang, Chong Luo, Bei Liu, Yiming Li, [0.58em] Tsinghua University Microsoft Research Asia MBZUAI [0.25em] https://github.com/THUSI-Lab/Causal-Reasoner |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 311 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.01810