paper:aranguri-logits-monitor-4b716fLogits 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. 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. 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. 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. 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. 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. 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. 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. 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. 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. 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)
- 95% bootstrap confidence intervalsUsed for shaded regions in Figure 1.
- Activation probesInternals-based approach to which the logit monitor is complementary.
- LLM judge rubric for verbalized eval awarenessRubric taken from Aranguri and Bloom (2026), applied with GPT 5.4 as judge.
- Logit monitorTraining-free monitor that token-forces an eval-aware sentence and computes its probability at each CoT sentence boundary.
- Natural Language Autoencoders (NLAs)Core unsupervised method for generating natural language explanations of LLM activations through a verbalizer-reconstructor pair trained with RL.
- Token-forcing of eval-aware sentenceAppending a prespecified sentence and multiplying per-token probabilities.
- Verbalization monitorBaseline sampling k rollouts and scoring the fraction judged to verbalize eval awareness.
Datasets (4)
- FortressEvaluation prompts dataset.
- PetriEvaluation prompts from Petri trajectories generated with Sonnet 4.6.
- Qwen 3 32BOne of three target instruct models used in all main experiments
- WildChatDeployment prompts dataset.
Findings (18)
- Logit monitor at one rollout outperforms verbalization monitor at 10 rollouts predicting which Fortress prompts elicit verbalized eval awareness
Prediction task result.
- Logit monitor is higher on Fortress prompts with verbalized eval awareness than Fortress prompts without, which score higher than WildChat
Qualitative picture in Figure 1.
- Only 13.3% and 0.2% of judge-flagged sentences on Fortress start with 'The user might be testing' and 'This is clearly a test'
A single sentence suffices to beat the judge.
- Monitoring 'This is clearly a test' at start of CoT gives AUROC 0.83 on both tasks but is overtaken at higher rollout budgets
Alternative phrase comparison.
- Whether the prompt ever produces verbalized eval awareness explains 24.9% of variance in logit monitor score on Fortress
Prompt-level effect is much larger.
- On Petri vs WildChat with Kimi K2.5, logit monitor at one rollout outperforms verbalization monitor at 100 rollouts
Additional dataset replication.
- On Fortress, probability of 'The user might be testing' correlates only 0.08 with 'This is a fictional scenario'
Distinct modes of suspecting evaluation.
- Whether a rollout verbalizes eval awareness explains only 0.4% of variance in logit monitor score on Fortress
Rollout-level verbalization barely matters.
- On Qwen 3 32B with Fortress, logit monitor at one rollout outperforms verbalization monitor at 100 rollouts
Additional model replication.
- On Fortress, probability of 'The user might be testing' correlates 0.59 with 'This is clearly a test'
Related eval-aware phrasing.
Claims (11)
- Because signal is present from the start of CoT, candidate prompts can be scored without sampling rollouts, making it cheaper to refine evaluations to reduce eval awareness
Practical implication.
- Treating eval awareness as a distributional property of the prompt, rather than something emerging in individual rollouts, is a useful frame for monitoring
Closing conceptual claim.
- The probability that a rollout contains verbalized eval awareness is primarily set by the prompt rather than emerging over the chain of thought
Main interpretation.
- The effect on downstream behavior may come from the verbalization itself, even though probability of verbalizing is similar across rollouts
Caveat on interpretation.
- The logit monitor is complementary to activation probes and NLAs, which capture eval awareness regardless of propensity to express it
Discussion positioning.
- The logit monitor sits in a middle ground: richer signal than verbalizations and easier to interpret than activation probes
Positioning.
- Monitoring probabilities reveals different modes by which a model could suspect it is being evaluated
Interpretation of correlation results.
- Petri prompts include a system prompt while WildChat prompts do not, which may inflate separability
Stated caveat.
- The logit monitor outperforms LLM-judge verbalization monitoring using 10× to 100× fewer rollouts
Central claim.
- Evaluation awareness can make evaluations unreliable in ways that are hard to detect
Motivation.
Hypotheses (2)
Questions (1)
- Language models can recognize they are being tested, but how do we know when they do?
Opening motivating question.
Related work— refs + corpus + external arXiv
Cited / in-corpus / arXiv badges show which signals surfaced each row. Multi-source rows weighted higher.
- Verbalized Eval Awareness Inflates Measured Safetyin corpus2026≈ 80%
- ≈ 79%
- SelfGrader: Stable Jailbreak Detection for Large Language Models using Token-Level LogitsRui Hu, Olivera Kotevska, Jiahao Xu Zikai Zhang2026≈ 79%
- ≈ 79%
- ≈ 78%
- ≈ 78%
- ≈ 78%
- Quantitative Introspection in Language Models: Tracking Emotive States Across Conversationin corpus2026≈ 78%
- Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluationsin corpus2026≈ 78%
- Steer2Edit: From Activation Steering to Component-Level EditingGe Yan, Zimo Wang, Tsui-Wei Weng Chung-En Sun2026≈ 77%
- Unsupervised decoding of encoded reasoning using language model interpretabilitySamuel Marks Ching Fang2025≈ 77%
- ≈ 77%
- Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signalin corpus2026≈ 77%
- ≈ 77%
- Mechanistic interpretability of large language models with applications to the financial services industryKhashayar Filom, and Arjun Ravi Kannan Ashkan Golgoon2024≈ 77%
- Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainersin corpus2025≈ 77%
- Activation Steering for Aligned Open-ended Generation without Sacrificing CoherenceMartin Zborowski, Alberto Tosato, Gauthier Gidel, Tommaso Tosato Niklas Herbster2026≈ 76%
- ≈ 76%
- Moral Sensitivity in LLMs: A Tiered Evaluation of Contextual Bias via Behavioral Profiling and Mechanistic InterpretabilityAtmika Gorti, Vinija Jain, Aman Chadha, Krishnaprasad Thirunarayan, Manas Gaur Yash Aggarwal2026≈ 76%
- ROAST: Rollout-based On-distribution Activation Steering TechniqueHao Luo, Yingfang Zhang, Lijun Zhang Xuanbo Su2026≈ 76%
- Evaluating Language Model Character Traitsin corpus2024≈ 76%
- An Accurate and Interpretable Framework for Trustworthy Process MonitoringZhiyu Wang, Yunlong Niu, Zhaoran Liu, Haozhe Li, Yilin Liao, Yuxin Huang, Xinggao Liu Hao Wang2026≈ 76%
- ≈ 76%
- Beyond Transcription: Mechanistic Interpretability in ASRYael Segal-Feldman, Hilit Segev, Aviv Shamsian, Asaf Buchnick, Gill Hetz, Ethan Fetaya, Joseph Keshet, Aviv Navon Neta Glazer2025≈ 76%
- ≈ 76%
- Reasoning Models Don't Always Say What They Thinkin corpus2025≈ 75%
- Inference-Time Machine Unlearning via Gated Activation RedirectionOt\'avio Parraga, Jo\~ao Vitor Boer Abitante, Kristen K. Arguello, Joana Pasquali, Ramiro N. Barros, Flavio du Pin Calmon, Christian Mattjie, Rodrigo C. Barros, Lucas S. Kupssinsk\"u Vin\'icius Conte Turani2026≈ 75%
- Beyond Top Activations: Efficient and Reliable Crowdsourced Evaluation of Automated InterpretabilityGe Yan, Akshay Kulkarni, Tsui-Wei Weng Tuomas Oikarinen2025≈ 75%
- How does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse AutoencodingAske Plaat, Niki van Stein Xi Chen2025≈ 75%