LLM Reasoning as Trajectories: Step-Specific Representation Geometry and Correctness Signals
Lihao Sun, Hang Dong, Bo Qiao, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan · 2026
Read LLM Reasoning as Trajectories: Step-Specific Representation Geometry and Correctness Signals by Lihao Sun, Hang Dong, Bo Qiao, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan (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-2604-05655.txt.
Description
Sun and five colleagues map language-model chain-of-thought as a trajectory through representation space. Mathematical reasoning crosses ordered, step-specific subspaces that sharpen with depth, and correct paths diverge from wrong ones only late. That late split lets mid-reasoning prediction of final correctness reach 0.87 ROC-AUC, and trajectory steering enables correction and length control.
Catalog record
| CREATOR | Lihao Sun, Hang Dong, Bo Qiao, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 51481 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.05655