Overthinking (in Reasoning Models)

Following the release of OpenAI o1 and DeepSeek-R1, RLVR training rewarded models for generating long chains of thought containing self-verification and backtracking steps. In January–March 2025, independent research groups—including Chen et al. ("Do NOT Think That Much for 2+3=?") and Cuadron et al. at UC Berkeley ("The Danger of Overthinking")—documented that reasoning models frequently determine the correct answer within the first few hundred tokens but continue generating thousands of tokens of self-doubt, circular checks, or over-complicated hypotheses in both math problems and software-engineering agents.

Sources: s1, s2, s3

Unbounded reasoning traces multiply inference latency and token costs by 3x to 10x on simple queries where standard instruction-tuned models succeed immediately. Worse, in agentic environments and abductive or visual reasoning tasks, prolonged internal monologue can cause the model to talk itself out of a valid initial solution, ignore environment feedback, or hallucinate nonexistent edge cases.

Sources: s1, s2

When asked a straightforward arithmetic or single-file bug-fix question, an uncalibrated reasoning model outputs the right solution in step two, but then emits twelve additional paragraphs beginning with "Wait, let me double check whether the problem statement could be interpreted in base 8..." before timing out or reverting to an incorrect edge-case patch. Applying a thinking-budget cap (`reasoning_effort="low"`) or a length-penalized RL reward eliminates the redundant verification loop.

Sources: s1, s3

Aggressive length penalties or hard token truncation during training and inference can backfire on genuinely hard multi-step problems where deep backtracking is necessary, requiring difficulty-aware routing rather than a single global token cap.

Sources: s2, s3