A new technique lets AI reasoning models decide, on their own, when a long chain of thought is overkill.
Researchers behind a paper called QA-Merging combine a Long-CoT model, trained to reason through lengthy step-by-step chains, and a Short-CoT model, trained to answer quickly, into one query-adaptive reasoner. Rather than retraining from scratch or leaning on prompt tricks, the method builds a small calibration set that labels which queries call for long or short reasoning. It then locates the specific transformer layers where long and short reasoning behavior diverges most and recalibrates only those layers through feature alignment and contrastive shaping, applying a simpler closed-form correction to the rest. The approach was tested on seven reasoning benchmarks across two model scales.
Long chains of thought are one of the biggest hidden costs of "thinking" models - every extra reasoning token adds latency and compute, even for questions that do not need it. Most existing fixes require retraining a model or brittle prompting that breaks when the wording changes. Treating reasoning style as a property of specific layers, rather than the whole network, is a cheaper middle path between those two extremes.
It is still a benchmark paper, not a shipping product - the real test is whether the trick holds up on the messy, ambiguous queries production systems actually see.