After DeepSeek-R1, the industry discovered a counter-intuitive fact: LRM writes hundreds of tokens of reflection even on "1+1=?", with token and latency being seriously wasted. CAT (Confidence-Adaptive Thinking), included in arXiv 2607.00862 and accepted by ACL 2026 Industry Track, treats the model's "self-confidence" as a governor, letting the LRM decide for itself how long to think on each problem. Previously, reasoning compression was done either by "blanket" CoT compression, or by attaching an external classifier to roughly judge difficulty — the former loses accuracy on hard problems, the latter is easily misjudged. CAT's observation is simple: the model actually "knows" how confident it is on each problem, and this self-certainty has rarely been used in preference optimization. The authors pour confidence as a preference signal directly into preference optimization, letting the model simultaneously learn two things: compress answers on confident problems, fully reason on unconfident problems. The same set of weights, two or three lines on "9.9 vs 9.11", carefully reason dozens of steps on proof problems. Stable SOTA on multiple benchmarks, with significant reduction in average token consumption. CAT's real contribution isn't a new algorithm, but pointing out a neglected free signal — the model's confidence in its own output naturally varies with the problem. When the industry changes RL objectives to "length-adaptive user needs", this "introspective" signal is more robust and cheaper than an external classifier, and will further spread in agent planning, multi-turn reasoning, and tool-calling scenarios.