The monitor that goes silent when AI reasoning fails

Low-bit quantization makes small reasoning models cheap to deploy, but it degrades their chain-of-thought reasoning. A natural instinct is to monitor the decoder and intervene when generation goes wrong. But a new study from Novelis Research argues that the most obvious observable, token log-probability, is the wrong basis for such a monitor.

Why log-probability fails

The study, led by El Hassane Ettifouri and colleagues, shows that centering a token log-probability under the model's own sampling law produces a valid martingale. But that martingale measures self-consistency of sampling, not trajectory health. When a model enters a confident repetition loop, the token has probability near 1, log-probability near 0, and entropy near 0. The monitor goes silent just as the model goes confidently wrong. This blind spot echoes findings on chain-of-thought monitoring: frontier models can only control their thinking 2.7% of the time, per a parallel study on safety gaps.

In a controlled sanity check, the log-probability martingale raised alarms on 63% of healthy streams, a false-alarm rate that disqualifies it as a detector, yet detected confident loops only 42% of the time. The calibrated e-CUSUM detector caught 100% of confident loops and drift-then-collapse regimes without false alarms on healthy text.

The calibrated e-CUSUM replacement

The researchers built a training-free decoder controller out of two pieces. First, a degeneration-aware alarm score that fuses token uncertainty with an explicit verbatim-repetition signal. Second, a calibrated e-process-based sequential detector that converts the raw product process into a CUSUM-floored statistic for change detection.

Calibration was decisive. An uncalibrated baseline (μ0=0.15) caused the detector to fire on 93-95% of all generations, making it useless. Setting μ0 at the 90th percentile of healthy FP16 traces (0.41) turned it into a selective detector reaching precision around 0.6 and phi around 0.3, against a 0.38 base rate of bad traces. The need for careful calibration of monitoring tools mirrors what researchers found in dataset distillation; see M3D's approach to high-resolution efficiency.

Empirical findings on GSM8K

Using DeepSeek-R1-Distill-Qwen-1.5B on 100 GSM8K test problems, the study found that verbatim loops are rare, at most 4% of tokens even on failing traces. The dominant failure mode is non-termination: up to 49% of incorrect INT4 traces exhaust the 2048-token budget. This matches findings from Lotfi et al. (2026) that quantized models over-produce markers like 'wait' and 'but' at high-entropy positions. The non-termination failure is reminiscent of the verification horizon problem described in the verification horizon study.

The controller reduced verbatim-degeneration signals. INT4 mean consecutive repetition fell from 0.010 to 0.003, severe loops from 1% to 0%, and truncation from 22% to 19%. Accuracy moved in the right direction, plus 6 points in INT4 and plus 4 in FP16, but the differences were not statistically significant (McNemar p=0.18 for INT4 at n=100). The token cost was plus 28%, and wall-clock time increased by 30%.

Methodological implications

The study's lasting contribution is methodological: a clear account of why a tempting observable is inadequate, plus a calibrated, honestly-evaluated replacement. The raw product process has a genuine Ville-style time-uniform false-alarm guarantee under a conditional null, but the deployed CUSUM statistic is reported through empirical operating points rather than as a strict mathematical guarantee. This honest accounting of limitations is something the field needs more of, as noted in an analysis of sandbox benchmark failures.

The researchers release all code, calibration tools, and traces to support larger-scale studies that the accuracy question ultimately needs. They recommend future work focus on harder benchmarks where non-termination and drift are more frequent, and on ablations that separate the contribution of entropy-only versus repetition-only signals in the alarm score. For those building similar monitoring pipelines, the principle is the same as in AI search infrastructure: get the plumbing right first, as the unified search framework shows.