Answers are grounded in this paper and cite their sources.
Reusing an FP32-calibrated conformal threshold on a different numerical implementation can change individual prediction sets even when aggregate coverage changes little. We evaluate source-side input rankings for finding these disagreements with a limited number of target executions. The methods include threshold-margin ranking (PG-M), a weight-change sensitivity ranking (PG-R), confidence-based controls and a DFLARE-style input generator. The protocol was developed on 30 historical MLPs and evaluated once on fresh CIFAR-10 models: five seeds each of a small CNN and ResNet-18, across PyTorch and ONNX Runtime backends. With 100 target queries per model, PG-M found all 36 CUDA BF16 and all 7 FP16 disagreements of the small CNN, 129 of 136 INT8 disagreements, and all 75 INT8 disagreements of ResNet-18. The largest expected count under uniform selection was 5.4 (small CNN, INT8). For the probability score on the test set, mean coverage changed by at most 0.08 percentage points, while up to 565 prediction sets per backend changed, summed across five seeds at α = 0.1. The sensitivity term provided little additional benefit, and confidence-based controls were less effective at small query budgets. In a separate evaluation of explicit floating-point MLP programs, certificate screening reduced target calls by 95–99.5% but took 35–130 times longer than PG-M.
No takes yet. Share an insight, caveat, or question.
S. M. Abdulla Al Mamun (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: