Friction-Guided Inference: Calibrating Correction Strategies and Abstention from Logprob Signals

Large language models frequently possess the knowledge needed to answer a question correctly yet commit to the wrong response. This paper presents friction-guided inference, a calibrated inference-time pipeline that uses the model's own logprob distribution — available at zero cost from any OpenAI-compatible API — both to select calibrated correction strategies and to identify questions where the model should abstain rather than commit. The key signal is competing routes (CR): the count of high-probability alternative tokens per position. CR detects when a model is uncertain and, through calibration, identifies which correction strategies help; an ablation shows CR does not determine which answer is correct at the individual-question level, so the lift comes from the strategy itself combined with logprob-derived uncertainty thresholds. On held-out data the calibrated strategy pipeline produces +7.7 to +20.8 pp across four model–benchmark cells (mean +11.8 pp, all statistically significant), and entropy-guided abstention alone produces +9.0 to +15.7 pp success-rate improvement at 20% abstention, at zero additional inference cost. The four cells span three architecture families (dense transformer, mixture-of-experts, Liquid Neural Network) and three benchmarks (MATH-500, SimpleQA, MMLU-Pro). The method requires per-model, per-benchmark calibration; the pipeline code and signal extraction transfer across the architectures tested. v4 update (July 2026). The empirical results are unchanged. This version strengthens the paper's positioning against prior work. First, the abstention component is now stated plainly as established prior art that the pipeline re-uses rather than contributes: thresholding a confidence signal derived from a model's token distribution to decide when to abstain is a mature technique, and the related work now cites that field (early-abstention cascades; abstention benchmarks; trajectory- and incentive-based abstention learning). The paper's own contribution in that area is clarified as the empirical analysis — that plain entropy outperforms the competing-routes count as an abstention signal, and the negative case where no such signal detects errors. Second, a new related-work discussion names the formal lineage of the pipeline's core operation — deciding how much computation to spend from an internal signal — in the adaptive-computation-time, early-exit, and rational-metareasoning literatures, and distinguishes this work by its whole-question granularity, its use of a frozen model through a public interface, and its finding that a signal good enough to gate computation need not be good enough to select among outputs. Several bibliographic author attributions are corrected. Builds on v3.