Stable Answers, Unfinished Reasoning: Why Self-Consensus Is Not a Safe Early-Exit Signal

Researchers investigating methods to lower the inference costs of reasoning models have found that using self-consensus—stopping once repeated probes yield the same answer—is not a reliable indicator of accuracy. The study suggests that models can frequently reach a stable, incorrect consensus, meaning this technique may provide a false sense of security while failing to save tokens effectively. This finding highlights a significant trade-off between computational efficiency and reliability in current automated reasoning systems.