Detectability of Doomed Reasoning Chains

01

Research question

Can a model’s internal representations reveal that a reasoning trajectory is unlikely to succeed before the model finishes generating it? If that signal is reliable, it could help allocate rollout compute more selectively during RLVR-style training.

02

Experimental setup

I generated approximately 16,000 GRPO-style rollouts from DeepSeek-R1-Distill-Qwen-1.5B on MATH problems, graded them with a verifier, and extracted residual-stream hidden states along each reasoning chain. Linear probes were trained to predict the eventual binary outcome from intermediate representations.

Probe performance was evaluated both in aggregate and by relative position within the generated chain. Threshold sweeps then simulated early termination and measured the tradeoff between token savings and preservation of the original GRPO advantage signal.

03

My contribution

I initiated and lead this independent study. I designed the question and experimental protocol, generated and verified the rollout dataset, implemented hidden-state extraction and probe training, and built the truncation and advantage-fidelity analyses.

04

Current results

Across pooled hidden states, probes reached approximately 0.708 AUC. Predictive performance rose from roughly 0.63 early in a chain to 0.80 later, consistent with outcome information becoming more legible as reasoning unfolds.

At one simulated operating point, 80% kill precision produced 7.4% token savings while preserving a 0.93 correlation with the untruncated GRPO advantages. A more aggressive policy saved approximately 23% of tokens, but advantage correlation fell to roughly 0.75.

05

Limitations & next questions

The strongest caveat is difficulty-versus-trajectory confounding: a probe may identify hard problems rather than uniquely diagnosing a doomed attempt. Problem-level splits and within-group analyses expose this issue, but do not fully resolve it.

Calibration across problem types, uncertainty around threshold selection, statistical validation, and generalization to other models and domains remain open. The current evidence supports further investigation—not a claim that probe-guided truncation is ready for training use.