Jul 2026
Calibrated e-CUSUM Decoding for Quantized Reasoning Models: Why Token Log-Probability Is the Wrong Observable for Decoding Monitors
A training-free decoding controller that combines a degeneration-aware alarm score fusing token uncertainty with explicit verbatim repetition and a calibrated e-process-inspired sequential detector, and finds that non-termination, rather than looping, is the dominant failure mode on GSM8K.
E. Ettifouri, Ayoub Belfatmi, Mahaman Sanoussi Yahaya Alassan et al.
· arXiv.org · 0 citations