Token Budget Saturation and Mechanistic Early Detection of Reasoning Non-Convergence in Chain-of-Thought Models
Non-converged reasoning runs were almost always wrong, and the authors found a weak early signal inside the model before the budget ran out.
The paper studies DeepSeek-R1-Distill-Qwen-7B on AIME problems and reports a split between chains that finish within budget and chains that exhaust it. Finished generations reached 90.3% accuracy; non-converged ones reached 6.6%, with 62.0% converging overall. Linear probes on hidden states, especially layer 20 around token 150, predicted convergence modestly above chance and beat entropy or repetition baselines. The authors say the signal is promising but not conventionally confirmed by their sweep-level permutation test. ArXiv · AI/CL/LG's note
The paper studies DeepSeek-R1-Distill-Qwen-7B on AIME problems and reports a split between chains that finish within budget and chains that exhaust it. Finished generations reached 90.3% accuracy; non-converged ones reached 6.6%, with 62.0% converging overall. Linear probes on hidden states, especially layer 20 around token 150, predicted convergence modestly above chance and beat entropy or repetition baselines. The authors say the signal is promising but not conventionally confirmed by their sweep-level permutation test. ArXiv · AI/CL/LG's note
score 5