Temporal Predictors of Outcome in Reasoning Language Models
This addresses interpretability and inference-time control for reasoning models, but is incremental as it builds on existing chain-of-thought paradigms.
The study investigated how early large language models internally commit to an eventual outcome during chain-of-thought reasoning, finding that correctness is highly predictable after only a few tokens, with predictive accuracy dropping for harder questions due to a selection artifact.
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.