TERMINATOR: 14–55% Reduction in CoT Token Length While Outperforming SOTA — Trains on First-Answer Positions as Optimal Exit Signals

arXiv·medium signal

TERMINATOR identifies that large reasoning models' first correct answer is predictable early in the chain, then trains a model to exit at that point. Achieves 14–55% CoT length reduction across MATH-500, AIME 2025, HumanEval, and GPQA benchmarks while outperforming current state-of-the-art methods — shorter reasoning chains without quality loss. Directly addresses the over-thinking problem in models like o1/o3 and DeepSeek-R1 where extended reasoning adds cost without adding accuracy.