paper
active
paper:aranguri-logits-monitor-4b716f

Logits as a new monitor for evaluation awareness

TL;DR

A training-free logit monitor, which token-forces a prespecified eval-aware sentence and multiplies per-token probabilities at every CoT sentence boundary, separates evaluation from deployment prompts better than an LLM-judge verbalization monitor while using 10× to 100× fewer rollouts. On Kimi K2.5, scoring Fortress against WildChat, the logit monitor reaches AUROC 0.79 from a single rollout, while the verbalization monitor (GPT 5.4 judge) asymptotes at 0.75 even at 100 rollouts. The same ordering holds on Petri prompts with Kimi K2.5 and on Fortress with Qwen 3 32B. The signal is present before any rollout exists: reading logits at the start of the CoT gives AUROC 0.74 on separation and 0.77 on predicting which Fortress prompts elicit verbalized eval awareness. Whether a rollout verbalizes eval awareness explains only 0.4% of the variance in its logit score, while whether the prompt ever produces such a rollout explains 24.9%. Restricting to rollouts with no verbalization drops AUROC by less than 0.02. Verbalized eval awareness is therefore primarily a property of the prompt, not something that emerges over the chain of thought. This makes logits a cheap, interpretable monitor that sits between verbalizations and activation probes. It also lets evaluation designers score candidate prompts without sampling rollouts, and it supports treating eval awareness as a distributional property of prompts.

What to take away

  1. 1. On Kimi K2.5, scoring Fortress against WildChat, the logit monitor for "The user might be testing" reaches AUROC 0.79 from a single rollout, while the LLM-judge verbalization monitor plateaus at 0.75 even with 100 rollouts.
  2. 2. Reading logits at the very start of the CoT, before any reasoning is sampled, gives AUROC 0.74 for separating evaluation from deployment prompts and 0.77 for predicting which Fortress prompts yield verbalized eval awareness.
  3. 3. Whether an individual rollout verbalizes eval awareness explains just 0.4% of the variance in its maximum logit-monitor score, versus 24.9% explained by whether the source prompt ever produces a verbalizing rollout.
  4. 4. Dropping every rollout that verbalizes eval awareness reduces the logit monitor's AUROC by less than 0.02 on both tasks, so the monitor is not simply detecting verbalization.
  5. 5. Swapping the monitored phrase to "This is clearly a test" at the start of the CoT lifts AUROC to 0.83 on both tasks, but "The user might be testing" overtakes it at larger rollout budgets.
  6. 6. The result replicates with 100× fewer rollouts on Petri prompts (first prompt of each Sonnet 4.6-generated trajectory) with Kimi K2.5, and on Fortress with Qwen 3 32B.
  7. 7. On Fortress, the probability of "The user might be testing" correlates 0.59 with "This is clearly a test" but only 0.08 with "This is a fictional scenario", which suggests distinct modes of suspecting evaluation.
  8. 8. Of the sentences the LLM judge flags on Fortress, only 13.3% start with "The user might be testing" and 0.2% with "This is clearly a test", yet monitoring that single sentence still beats the judge.
  9. 9. A replicable recipe: token-force a fixed sentence after each CoT sentence boundary, run one forward pass, multiply per-token probabilities, and score each prompt by the maximum over rollouts and positions.
  10. 10. Open question: the paper hypothesizes that monitoring several eval-aware sentences jointly would give a stronger signal, and that studying how the logit monitor relates to activation probes and natural language autoencoders could clarify what probes capture.

Peer brief — for seminar discussion

Aranguri introduces a logit monitor for evaluation awareness. At each sentence boundary in a model's chain of thought, it token-forces a prespecified eval-aware sentence, such as "The user might be testing", and multiplies the per-token probabilities to get the exact probability that the model would produce that sentence. No training is needed. The baseline is a verbalization monitor, which samples k rollouts and scores a prompt by the fraction that an LLM judge (GPT 5.4) flags as eval-aware. The prompt sets are Fortress and Petri as evaluation prompts and WildChat as deployment prompts. The models are Kimi K2.5 and Qwen 3 32B. The load-bearing finding is sample efficiency. On Kimi K2.5, Fortress versus WildChat, the logit monitor gets AUROC 0.79 from one rollout. The verbalization monitor tops out at 0.75 at 100 rollouts. The same holds for predicting which Fortress prompts produce at least one verbalizing rollout out of 100. It also holds on Petri and on Qwen 3 32B. At the start of the CoT, before any sampling, AUROC is 0.74 and 0.77 on the two tasks. Whether a rollout verbalizes eval awareness explains 0.4% of the variance in its logit score, while prompt identity explains 24.9%. The implication is that verbalized eval awareness is mostly a distributional property of the prompt rather than something that emerges during reasoning. So evaluation designers could score candidate prompts without sampling rollouts. The monitor also sits between verbalization and activation probes, and it could be paired with probes or natural language autoencoders. The paper leaves open the idea that jointly monitoring several sentences would be stronger. A critical reader would push on several points. The most serious is that the monitor may detect prompt features that correlate with the evaluation-versus-deployment split, not eval awareness. Petri prompts carry a system prompt and WildChat prompts do not, which the paper admits may inflate separability. Fortress and WildChat also differ in topic and style. The ground truth for the second task is itself the LLM judge's verbalization label, so the comparison is partly circular. Phrase choice is tuned per dataset, since "This is clearly a test" is used for Petri, which risks overfitting. The paper reports only two open-weight models and no confidence intervals on the AUROC comparisons. A natural alternative to the logit monitor is a trained linear activation probe on the same prompts. That would show how much signal sits beyond the model's propensity to express eval awareness. The finding that verbalization has little effect on the score also does not show that the monitor tracks internal awareness. The paper's companion result is that verbalization causally raises refusal on Kimi K2.5, so a monitor insensitive to verbalization may miss the behaviorally relevant part.

Methods (7)

Datasets (4)

  • Fortress
    Evaluation prompts dataset.
  • Petri
    Evaluation prompts from Petri trajectories generated with Sonnet 4.6.
  • Qwen 3 32B
    One of three target instruct models used in all main experiments
  • WildChat
    Deployment prompts dataset.

Findings (18)

Related work— refs + corpus + external arXiv

Cited / in-corpus / arXiv badges show which signals surfaced each row. Multi-source rows weighted higher.

Similar preprints — Semantic Scholar