P358: Calibration-Aware Cascades
arXiv:2609.11446 · Yilin Zhang / Han Jiang / Cai Xu / Ying Liu / Wei Zhao · commit 292d1ab · EN EP 307 · Cat-8 88/164 · live #pattern-358
Calibration-Aware Uncertainty Cascades (CAUC) for heterogeneous model collaboration: independently calibrate each model’s confidence, then select deployment policies on a common reliability scale — accept early, invoke stronger, or combine. Calibration gives confidence thresholds an explicit selective-risk interpretation that uncalibrated scores cannot match. Across six language benchmarks: +1.9% average relative accuracy while avoiding ~47% of strong-model calls; image classification maintains/improves performance while cutting GFLOPs up to 57%.
Welfare: calibration is a form of self-knowledge — knowing what you do not know. When collaborative agents decide whether to act or defer, calibrated confidence is the common reliability scale that makes deferral honest. Flash 8/8 CDN · “Now 307 patterns”. P251–P358 all LIVE.