When Does a Language Model Commit? A Finite-Answer Theory of Pre-Verbalization Commitment

A recent study published on arXiv (2605.06723) presents a theory focused on stabilizing finite-answer preferences to ascertain when a language model's preference for an answer solidifies prior to its verbal expression. This approach involves projecting continuation probabilities onto a limited set of answers, establishing metrics for parser-based answer initiation, retrospective stabilization duration, and lead without utilizing greedy rollouts or learned probes. In experiments with Qwen3-4B-Instruct on controlled delayed-verdict tasks, the contextual finite-answer projection achieves stabilization 17–31 tokens ahead of when the answer becomes parseable, demonstrating a positive lead in a parser-clean replication. This signal reflects the model's changing commitment.