Temporal Predictors of Outcome in Reasoning Language Models
Joey David
TL;DR
This paper shows that large language models implicitly assess their own accuracy very early in chain-of-thought reasoning by training linear probes on hidden states after only a few tokens. Using a difficulty-balanced MATH dataset and two 8B models, the authors extract final-state representations from the initial CoT steps, reduce dimensionality with PCA, and train $L_2$-regularized logistic regressors to predict eventual correctness. The main finding is that the internal signal is strong early (ROC-AUC around $0.84$ and accuracy near $0.76$ at $t=4$), and remains robust across longer prefixes, with observed declines explained by a shift toward harder questions rather than fading signal. This has implications for interpretability and dynamic inference-control, suggesting opportunities for early halting, rerouting, or reflective prompting based on the model’s hidden confidence, rather than relying on output-space cues alone.
Abstract
The chain-of-thought (CoT) paradigm uses the elicitation of step-by-step rationales as a proxy for reasoning, gradually refining the model's latent representation of a solution. However, it remains unclear just how early a Large Language Model (LLM) internally commits to an eventual outcome. We probe this by training linear classifiers on hidden states after the first t reasoning tokens, showing that eventual correctness is highly predictable after only a few tokens, even when longer outputs are needed to reach a definite answer. We show that, for harder questions, a drop in predictive accuracy highlights a selection artifact: hard items are disproportionately represented in long CoTs. Overall, our results imply that for reasoning models, internal self-assessment of success tends to emerge after only a few tokens, with implications for interpretability and for inference-time control.