Calibrated e-CUSUM Decoding for Quantized Reasoning Models: Why Token Log-Probability Is the Wrong Observable for Decoding Monitors

Researchers demonstrate that token log-probability is an ineffective metric for monitoring decoding quality in quantized reasoning models and propose an alternative calibration method.

Firefly links to the original publisher. The summary above is AI-generated for orientation and may differ from the source. The “current rank score” decays over time so newer significant stories surface first.