Overthinking in Language Models

This field studies why language models sometimes keep reasoning after reaching a useful answer, repeating steps, exploring unproductive branches, or even replacing a correct answer with a wrong one. Such behavior wastes tokens and latency in tasks like math, coding, and science question answering, and can also make agent behavior vulnerable to prompts or tool calls that trigger loops. Approaches include early stopping based on confidence, answer stability, or reasoning progress; difficulty-aware compute budgets; and training or inference-time methods that prune redundant steps or steer the model out of loops. The central tension is that useful deliberation and self-correction can look like waste, so reducing length must not sacrifice accuracy, faithfulness, or the ability to handle hard problems.