LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning

Large reasoning models achieve strong performance by generating long chains of thought, yet extended reasoning is often counterproductive: models frequently continue past the point of sufficiency, wasting compute and sometimes degrading correctness by revising a correct solution into an incorrect one. We show that reasoning models encode a domain-general \emph{readiness} signal in their hidden states, concentrated at natural high-uncertainty points such as "wait" and "hmm", that predicts whether stopping at a given moment would yield the correct answer. Building on this finding, we propose LYNX, an online early-exit framework that reads out this signal during generation and wraps it in split conformal calibration, giving practitioners a single confidence dial with calibrated control over erroneous exits rather than per-task heuristic thresholds. The readout requires no external verifier, auxiliary model, or additional human annotation: stop/continue labels are obtained by forcing the base model to answer at candidate stopping points and comparing to the task answer. A single LYNX head trained and calibrated once on a generic mathematical corpus transfers unchanged across benchmarks, decoding temperatures, and non-mathematical domains including commonsense reasoning and code generation, suggesting that readiness reflects model-internal representations rather than task-specific heuristics. Across three model families spanning 1.5B--32B parameters, LYNX matches or improves baseline accuracy while reducing generated tokens by 30--70\%, with competitive or superior accuracy--efficiency Pareto frontiers relative to prior early-exit methods.