When it comes to grading a monitor that watches a language model while it generates, AUROC has quietly become the default answer. A set of experiments across three open models suggests the problem: a monitor selected because of its AUROC may only speak up after the failure is already fully in progress.

The work is now on arXiv. 'Detection Is Not Early Warning: Evaluating Token-Level Alarms on Language Model Failures' (https://arxiv.org/abs/2611.16803) comes from researchers at ETH Zurich, the ETH AI Center and Solaris Research, and was accepted at the NeurIPS 2026 Workshop on Interpretability as a Science. Its central claim is that many language model failures begin at one specific token, which the paper calls the onset, and that monitors are silently being asked to do three very different jobs at once.

Three tasks, flattened into one metric

The paper separates monitoring into three layers. Response classification gives a single verdict on a response that is already finished. Failure detection raises an alarm once the failure is under way. Early warning raises one before the failure begins, and only that last tier leaves any room to stop, steer or resample the generation.

The authors introduce a protocol they call Prewarn. It calibrates each monitor's threshold on validation data to an error rate chosen for the application, runs the monitor, and then scores every alarm relative to the onset. The accompanying metrics include Warn_256, the share of flagged positions in the 256-token window right before the onset; Early, the share of responses whose first alarm precedes the onset; and Lead, the median gap between first alarm and onset, negative when the alarm is a warning.

The same 0.999, very different behaviour

The sharpest contrast comes from the token-cap rule, which reads nothing but the length of the response. On Apertus its AUROC is 0.9990, on Llama 0.9988 and on Mistral 0.9995, the latter two being the highest response-level scores among the three models. In the same table, its early-warning rate is 0.000, its windowed flag coverage is 0.000, and its Lead reads +3276, +2958 and +3688 tokens. In other words it knows almost perfectly well that a response has already fallen apart, and knows nothing at all about when the falling apart began. Its alarm arrives more than three thousand tokens after the onset.

The paper pushes the construction further. Four scorers reach almost the same AUROC, between 0.9980 and 0.9998, yet their warning coverage differs several-fold. Shuffle the scores and warning coverage can reach 0.52 to 0.67 with an early-warning rate of 0.95 to 0.96; keep only the highest score of each response and those figures collapse to 0.000 and 0.01 to 0.02. Identical AUROC, very different warning ability.

The authors also report where response-level metrics hit their ceiling. AUROC stays above 0.96 at every layer and best accuracy varies by less than 3%, but warning coverage is seven times higher at the best layer than at the worst. Ranking by AUROC agrees with the test ranking by warning coverage only weakly, at tau=0.33, while ranking by validation warning coverage predicts it closely, at tau=0.94.

Probes can read the signal, but not the earliness

The case study uses linear probes on activations across Apertus-8B-Instruct, Llama-3.1-8B-Instruct and Mistral-7B-Instruct. The prompts come from seven sources spanning mathematics, code, instruction following and medical reasoning, with ten sampled responses per prompt, and degeneration labels produced by Claude Sonnet 5 and then re-checked by hand.

The results split by model. On Apertus and Llama the probes raise their first alarm before the onset on most degenerate responses, with the median first alarm landing 137 to 232 tokens ahead of the onset, and they flag a larger share of the warning window than of the tokens before it, by 1.4 to 3 times. Mistral behaves differently: every probe stays below one half on the early-warning rate, the median first alarm falls after the onset, and yet the detection rate is still 0.99.

Adding rank-16 LoRA adapters underneath the probe more than doubles warning coverage on every model, while barely moving the early-warning rate. The table makes this plain: on Apertus the frozen probe's windowed flag coverage goes from 0.186 to 0.523 once adapters are attached, while the early-warning rate moves only from 0.658 to 0.695. The authors are careful about the reading here, noting that most of the added flags land earlier and further from the onset.

Move to held-out sources and the capability shrinks further. At a 5% false-alarm budget, the frozen probes' early-warning rates on Llama and Mistral drop to 0.20 and 0.12.

The same flaw in a different field

The problem is not confined to degeneration detection. In a separate arXiv paper (https://arxiv.org/abs/2606.12476), Itkin frames hallucination onset detection as a quickest change detection problem and derives a Lorden lower bound: at a false-alarm rate of 0.01 the detection delay cannot go below about 1.3 tokens, while under a recall-honest accounting the actual delay runs 56 to 66 tokens. His conclusion lands on the same point as this paper: classification metrics conceal that delay structure, and only sequential analysis makes it measurable.

Kavumba and colleagues (https://arxiv.org/abs/2604.03962) attack it from the other side. Rather than detecting the boundary, their streaming guardrail forecasts how harmful the likely continuations will be. On the QWENGUARDTEST response_loc task it reaches 97.5 F1, 95.1 recall and 92.6% on-time intervention, against 95.9, 92.1 and 89.9% for the Qwen3Guard-Stream-8B-strict baseline. A further example cited in the paper points the same way: a sepsis prediction model reaches an AUROC of 0.87 over all its predictions, but 0.62 on predictions made before sepsis criteria are met and 0.47 on predictions made before clinical recognition.

What is worth copying

Prewarn is not a demanding protocol. Mark the onset, calibrate the threshold, score by timing rather than by the final score. The hard part is the annotation, and admitting that a reported AUROC of 0.999 never showed the monitor was early enough to act.

The code and the probes with calibrated thresholds are released (https://github.com/swiss-ai/degeneration-probe). Next time you look at an evaluation table for a monitor, check whether it has a column for lead time. If it only has AUROC and accuracy, something is probably being scored from the final whistle instead of from the moment the game turned.