AI reasoning models waste tokens thinking hard about easy questions — a new framework aims to fix that.
Researchers introduced Confidence-Adaptive Thinking (CAT), a framework that reads a model's own internal certainty signals to decide how long it should reason before answering. Simple queries get short chains of thought; harder ones get the full treatment. The approach folds these confidence signals into the model's preference optimization process, so the model learns to calibrate its own verbosity rather than relying on an external length budget. In benchmark tests, CAT outperformed existing compression baselines on reasoning accuracy across several models.
The problem CAT targets is real and getting more expensive: models like DeepSeek-R1 and OpenAI's o-series generate long internal monologues before every answer, which inflates inference costs even on trivial tasks. Flat compression schemes — just cut token budgets uniformly — tend to hurt performance on genuinely hard problems, creating an accuracy-cost tradeoff that prior work hasn't cleanly resolved. CAT's self-certainty approach tries to thread that needle by making length reduction conditional.
Whether "intrinsic self-certainty signals" reliably track actual problem difficulty at production scale is the question labs will need to answer — a model can be confidently wrong, and overconfident compression on the wrong query is exactly the failure mode this method is supposed to prevent.