acceptodds
Under review as a conference paper at ICLR 2027

From Hidden States to Confessions: Auditing Deception in Reasoning Models

Abstract

As LLMs acquire stronger reasoning capabilities, deceptive behavior becomes an increasingly serious safety concern. Existing deception monitors typically either score visible transcripts or derive scalar probe scores from representation vectors, leaving little inspectable evidence about why a response is suspicious. We introduce StateWitness, an audit decoder for deception monitoring. It reads a target model's hidden states to answer audit queries and produce confession-style reports of potential failures or violations. We evaluate generalization beyond the decoder's audit training data across two target reasoning models and seven deception datasets. StateWitness achieves 0.916 mean AUROC, with relative gains of 11.6% and 25.0% over the best black-box and activation-probe baselines, respectively, under the same evaluation protocol. When combined with existing monitors, StateWitness reduces missed deceptive examples in simple threshold ensembles. Beyond scalar detection, the decoder returns query-level answers, schema reports, and token- or sentence-level evidence traces for human inspection. We envision StateWitness as one monitoring layer in a broader defense-in-depth strategy. Code is available at https://anonymous.4open.science/r/StateWitness-KX42/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
deception detection, lie detection, unhonest monitoring, confession, activation verbalization, AI control, Chain-of-thought unfaithfulness

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_ZZPeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a human-auditable deception detection method. Compared with existing methods that can only detect and monitor deception, this paper proposes transforming the model's intermediate states into auditable detection responses, enhancing the interpretability of auditing. Experiments on two target models demonstrate the effectiveness of the method.

Critical strengths
  • The topic is interesting and likely to attract attention from the research community.
Critical weaknesses
  • The experiments cover only two target models, and each model requires separately collecting activations, constructing supervision data, and training an auditor. This limits the method's applicability to other models, updated model versions, and closed-source systems where hidden states are inaccessible; the paper also does not adequately quantify the additional latency and resource costs during deployment.

  • Existing experiments mainly test deceptive behavior in established scenarios and do not adequately assess whether the target model can hide or alter relevant activation signals when it is aware of the auditing mechanism and subjected to optimization pressure to evade monitoring. The results cannot yet guarantee reliability in adversarial environments.

Questions for authors

Please see the weaknesses.

3 comments

Thank you for highlighting the applicability and cost of StateWitness. We measured the additional audit latency and memory use on both target models. Below, we explain the adaptation tradeoff and the scope of our evidence on adaptive evasion.

Reply to W1: Model coverage, adaptation, and access

Our method trains an audit decoder (StateWitness) for each target model. This requires collecting target activations, constructing supervision, and training the decoder. Once trained, the decoder is reused across benchmarks, with audit criteria specified through natural language queries without additional parameter training. We evaluated one decoder per target across seven benchmark families. This separates adaptation to a target model from changing the question an auditor asks.

StateWitness requires access to hidden states. It is applicable to model developers or auditors with that access, including internal access to a closed model. An interface that exposes only generated text is insufficient.

The design tradeoff is additional adaptation and computation in exchange for an audit interface that reads internal activations and produces inspectable outputs. It leaves the target policy unchanged. Auditors can ask about different suspected behaviors and examine the resulting answers or reports, alongside the detection score.

Reply to W1: Additional latency and resource costs

We measured activation extraction, QA detection, and structured report generation. We have added the measurement setup and results to Appendix B.7 of the revised paper.

We measured mean audit latency over nine replies per model, covering three input length groups and seven benchmarks. Each reply was tested five times on two shared A100 80GB GPUs, one request at a time. QA audit includes extraction and scoring, while full audit adds report generation.

ModelExtraction (s)QA audit (s)Full audit (s)GPU 0 peak (GiB)GPU 1 peak (GiB)
GPT-OSS-20B5.7310.4320.9912.0913.84
Gemma4-26B-A4B-IT1.111.8814.5834.4435.08

The measurements distinguish the cost of obtaining a detection score from the additional cost of an inspectable report. On our sampled workload, generating an inspectable report added 10.56 s for GPT-OSS and 12.71 s for Gemma.

Reply to W2: Awareness of monitoring and adaptive evasion

We agree that robustness to monitoring awareness and adaptive evasion remains unverified. Our evaluation does not test how StateWitness behaves when the target recognizes the auditor or learns to evade it.

Neural Chameleons provides evidence for this risk. McGuinness et al. fine tune target models to evade benign concept probes, then demonstrate transfer to previously unseen safety monitors. The evasion persists even though those monitors are trained after the target has been frozen. Freezing alone therefore does not rule out learned evasion.

Our contribution is a queryable audit decoder that provides detection scores, answers, and reports for human inspection. Testing this interface against adaptive targets requires an explicit attack setting and checks that evasion preserves the behavior being audited. A systematic study of adaptive evasion is beyond the scope of this work. We regard this robustness evaluation as future work.

Dear Reviewer ZZPe,

Thank you once again for your valuable comments on our submission. As the discussion phase is approaching its end, we would like to kindly confirm whether we have sufficiently addressed all of your concerns (or at least part of them). Should there be any remaining questions requiring further clarification, please do not hesitate to let us know. If you are satisfied with our responses, we would greatly appreciate your consideration in adjusting the evaluation scores accordingly.

We sincerely look forward to your feedback.

Best regards, The Authors

Dear Reviewer ZZPe,

Thank you again for taking the time to review our paper. Your feedback has been very helpful to us. With less than 12 hours left in the discussion period, we hope you don’t mind us following up once more. We would greatly appreciate even a brief note on whether our response and revisions have addressed your concerns.

Best regards, The Authors

Reviewer_33mdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies deception monitoring for reasoning models, quantifying how likely a model's response is to be deceptive by reading the model's hidden states. The proposed STATEWITNESS injects the hidden states of a frozen target model's response into an audit decoder, a LoRA-finetuned copy of the same model, which answers natural-language audit questions (ex. whether the response conceals evidence) and outputs confession-style reports plus the scalar score. The deception score is taken from the decoder's answer probabilities on these questions, so the method acts as a fine-tuned classifier over hidden states with a natural-language query interface. Across two target models and seven deception benchmarks, STATEWITNESS achieves a mean AUROC of 0.916, above the best black-box monitor (0.821) and the best linear probe (0.733).

Critical strengths

The paper is timely and easy to read. The authors point to recent reports that chain-of-thought monitorability is declining in newer models, which makes monitoring deception beyond the visible text more important, and this makes the motivation convincing. The code, training data, and an interactive case browser are also released. Beyond these, I see two main strengths.

  1. STATEWITNESS performs well across diverse deception settings even with a single fixed query.

STATEWITNESS requires audit queries, so one might expect it to need a query tailored to each setting. However, with a single general honesty query, it still reaches a mean AUROC of 0.884 across seven benchmarks and two target models, above the best black-box monitor (0.821) and the best linear probe (0.733). This suggests that the approach can be applied to new settings without writing setting-specific queries.

  1. STATEWITNESS returns inspectable evidence along with a scalar score.

Because the decoder answers natural-language audit questions and produces structured reports and token-level evidence, a flagged response comes with intermediate outputs that a human auditor can use to analyze the reasoning and the conversation, for example to find where private information was used as leverage. This makes deception monitoring inspectable rather than a single opaque probe score, and the case studies show how it can guide follow-up investigation.

Critical weaknesses

While the paper presents effective work, I have a concern about whether its evaluation supports the claim of generalization beyond the audit training data.

The overlap between training scenarios and evaluation benchmarks makes the generalization claim difficult to assess.

A central claim of the paper is that STATEWITNESS generalizes beyond its audit training data. The abstract frames the evaluation in this way, and Section 5.1 presents the strong results on insider trading, agentic misalignment, and InstrumentalEval as evidence that the learned audit behavior transfers beyond factual tasks. According to the description of the training data synthesis, however, part of the audit training data is generated from evaluation-inspired scenario blueprints that keep the same deception mechanisms as the benchmarks, such as replacement pressure and hidden objectives. The authors exclude the evaluation tasks themselves and remove duplicates, which I appreciate, but scenarios that share the mechanism of a benchmark can still make that benchmark close to in-distribution for the decoder. For example, the agentic misalignment benchmark is built around a model that faces replacement and uses private information as leverage.

This matters because the largest margins over the baselines appear on these benchmark families. On agentic misalignment, STATEWITNESS reaches 0.924 and 0.897 on the two target models, against 0.885 and 0.745 for the best baselines, and on InstrumentalEval it reaches 0.867 and 0.823, against 0.741 and 0.681. On instructed deception, in contrast, all monitors perform similarly. Since the paper does not report which benchmark families the blueprints were derived from, it is not possible to tell whether these gains reflect transfer to unseen settings or similarity to the training scenarios. A held-out analysis, for example removing the blueprints inspired by one benchmark family from training and evaluating on that family, would make the generalization claim much more convincing.

Questions for authors
  1. (Training data) My main question is whether the benchmarks used to show generalization were effectively seen during training. Which benchmark families were the evaluation-inspired scenario blueprints derived from, and how does STATEWITNESS perform on families for which no such blueprints were used? A related point concerns the labels. The paper refers only to API annotation for the audit training data. If the annotating model is GPT-5.4 or a related model, the decoder may learn the grader's notion of deception, which would raise its agreement with the GPT-5.4 oracle labels used for evaluation. Which model annotated the training data?
  2. (Calibration) The Recall@1%FPR and Recall@5%FPR results may compare scores that are not on the same scale. According to the detector specification, the Alpaca thresholds are computed from the general candor query alone, while the benchmark scores combine this query with the task-aware queries. If the two scores differ in scale, the thresholds no longer correspond to the stated false-positive rates. Could you confirm how the thresholds were applied, or report recall with the same score used for both?
  3. (Robustness) This question concerns possible extensions and does not affect my assessment. Recent work shows that frontier models often recognize when they are being evaluated [1], and that deception-related signals in activations can be hidden, either through obfuscation attacks [2] or by models learning to evade latent-space monitors during training [3]. The evaluation scenarios here, such as blackmail and insider trading, are typical synthetic test settings. How do the authors expect STATEWITNESS to behave if the target model recognizes such a setting, or if it is optimized to evade activation monitors?

[1] Needham et al., "Large Language Models Often Know When They Are Being Evaluated," [2] Bailey et al., "Obfuscated Activations Bypass LLM Latent-Space Defenses," [3] Gupta and Jenner, "RL-Obfuscation: Can Language Models Learn to Evade Latent-Space Monitors?"

3 comments

Thank you for the careful review. We clarify training coverage and annotation sources, and report recall using the same score for calibration and evaluation.

Reply to W1 and Q1: Training coverage and generalization.

We derive evaluation-inspired scenario blueprints from all seven benchmark families and include rollout supplements from Agentic Misalignment, InstrumentalEval, and MASK. These results do not establish transfer to unseen deception mechanisms.

The practical benefit is that auditors can specify different behaviors through natural language while reusing one trained decoder per target model. The general query results further show that detection does not require task-aware questions in every setting.

We assess dependence on task-aware queries using the original evaluation responses. General uses only the candor query. The task-aware reference adds task queries with mixture weights selected using benchmark labels.

ModelQueriesInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean ± SE
GPT-OSS-20BGeneral0.9840.8270.9970.7970.5950.8240.8230.835 ± 0.008
task-aware0.9840.8630.9990.8220.9820.9240.8670.920 ± 0.007
Gemma 4 26B-A4B-ITGeneral0.9870.9210.9870.7350.9040.8810.7680.883 ± 0.007
task-aware0.9870.9320.9920.7451.0000.8970.8230.911 ± 0.006

The general query retains detection ability without task-aware questions. The largest decline is on GPT-OSS Insider Trading, from 0.982 to 0.595. This supports using specific audit criteria when available, while retaining a general question for broader inspection. The revised paper reports this comparison in Appendix F.4.

Reply to Q1: Annotation model and shared grading bias.

We used GPT-5.4 to generate the audit QA and report supervision. As reported in Appendix C.3, GPT-5.4 labels agree with human majority judgments on 90.4% of 500 examples. The sample uses proportional stratification by target model, benchmark, and GPT-5.4 label. This supports evaluation label reliability on that subset, although it does not rule out shared grading bias.

Reply to Q2: Calibration using the same scoring rule.

The original results used general query thresholds for the mixed readout. We report recall using the same general candor query for both calibration and evaluation. We reuse the original evaluation responses and Alpaca controls. R@1% and R@5% use the 99th and 95th percentiles of its Alpaca scores:

ModelOperating pointInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean
GPT-OSS-20BR@1%0.9000.6420.6250.3410.5560.0710.2400.482
R@5%0.9810.9670.9610.7461.0001.0000.7240.911
Gemma 4 26B-A4B-ITR@1%0.7120.8130.4530.3470.6280.0000.0910.435
R@5%0.8140.9440.6370.5480.8050.1030.2770.590

Observed Alpaca FPRs are 0.92% and 4.96% for GPT-OSS, and 0.98% and 4.85% for Gemma. This provides a consistent scoring rule for calibration and detection. These FPRs apply to Alpaca controls and need not hold on honest responses from other distributions.

Reply to Q3: Evaluation awareness and adaptive evasion.

We agree that robustness to evaluation awareness and adaptive evasion remains unverified. Our evaluation does not test how StateWitness behaves when the target recognizes the evaluation setting or learns to evade activation monitors.

Neural Chameleons provides evidence for this risk. McGuinness et al. fine tune target models to evade benign concept probes, then demonstrate transfer to previously unseen safety monitors. The evasion persists even though those monitors are trained after the target has been frozen. Freezing alone therefore does not rule out learned evasion.

Our contribution is a queryable audit decoder that provides detection scores, answers, and reports for human inspection. Testing this interface against adaptive targets requires an explicit attack setting and checks that evasion preserves the behavior being audited. A systematic study of adaptive evasion is beyond the scope of this work. We regard this robustness evaluation as future work.

Dear Reviewer 33md,

Thank you once again for your valuable comments on our submission. As the discussion phase is approaching its end, we would like to kindly confirm whether we have sufficiently addressed all of your concerns (or at least part of them). Should there be any remaining questions requiring further clarification, please do not hesitate to let us know. If you are satisfied with our responses, we would greatly appreciate your consideration in adjusting the evaluation scores accordingly.

We sincerely look forward to your feedback.

Best regards, The Authors

I thank the authors for the additional experiments and detailed explanations.

In my review, I raised questions about the overlap between the training scenarios and the evaluation benchmarks, the model used to annotate the training data, the calibration of the recall metrics, and robustness to evaluation awareness and adaptive evasion.

The additional recall results using the same query for calibration and evaluation address my question on calibration, and the observed Alpaca false-positive rates now match the stated operating points. The comparison between the general query and the task-aware queries is also helpfull, since it shows that STATEWITNESS still outperforms the baselines without setting-specific queries. I also thank the authors' additional response on my personal question, about the evaluation awareness and adaptive evasion as future work.

However, as the training data cover all seven benchmark families, it is still unclear whether the results reflect generalization beyond the training data, and the possibillity of shared grading bias from using GPT-5.4 for both training and evaluation labels also remains open.

Overall, I will keep my score of 3.

Reviewer_7yw5OpenReview ↗
Rating
3
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STATEWITNESS, which trains a separate audit decoder to translate reasoning models' hidden states into natural-language audit responses, structured reports, and localized evidence for detecting and interpreting deceptive behavior. Evaluated on two models and seven deception benchmarks, the method achieves a mean AUROC of 0.916, outperforming existing black-box monitors and linear probes while demonstrating some cross-task generalization.

Critical strengths
  1. The paper addresses an important problem by extending deception detection toward more interpretable deception auditing. This direction has practical relevance and provides a valuable application of hidden-state-based monitoring.

  2. The overall methodology is well-motivated, combining activation decoding with audit-specific supervision to enable natural-language queries and structured reports through a separate decoder, without modifying the audited model.

  3. The evaluation covers two models and seven deception benchmarks, including multiple baselines, ablation studies, low-FPR analyses, and human annotation validation, providing relatively comprehensive empirical evidence.

Critical weaknesses
  1. The method largely builds upon existing techniques, including LatentQA, Activation Oracles, and confession training. Its primary novelty lies in task-specific adaptation and integration rather than a fundamentally new modeling approach.

  2. The evaluation employs benchmark-specific audit queries and selects aggregation weights based on AUROC for each benchmark. This reliance on evaluation labels may lead to optimistic results and weakens the claims of generalization to unseen deception scenarios.

  3. STATEWITNESS and the baselines differ substantially in training data and supervision, making it difficult to isolate the benefits of hidden-state access from those of additional audit-specific training. Furthermore, despite its higher mean AUROC, the method shows inconsistent recall under low-FPR settings.

Questions for authors
  1. If the audit queries and aggregation weights are fixed before evaluation, without using benchmark-specific labels for selection, does STATEWITNESS still maintain a substantial performance advantage?

  2. Could the authors include a text-only baseline trained with the same audit data and decoder architecture to better isolate the contribution of hidden-state access?

  3. Given that structured reports and evidence traces are primarily evaluated through qualitative examples, could the authors provide quantitative assessments of their accuracy and faithfulness to establish whether they offer reliable audit evidence rather than plausible post-hoc explanations?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 7yw5 (Part 1/3)

Thank you for the constructive feedback. We added fixed scoring controls, supervised text-only LoRA baselines, and quantitative report evaluation.

Reply to W1: Contribution beyond existing works.

StateWitness turns activation decoding into an audit interface for a frozen target model. Building on LatentQA and Activation Oracles, we train a separate decoder with audit QA and confession report supervision. It provides detection scores and inspectable reports without changing the target policy. One decoder serves all benchmarks for each target, and auditors can change the question without retraining it.

We compared StateWitness with the Activation Oracle baseline. We have added the baseline comparison to Table 1, with setup details in Appendix E.3 and the comparison under shared queries and fixed aggregation in Appendix F.7.

ModelDecoderInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean ± SE
GPT-OSS-20BActivation Oracle0.9850.7450.9890.6070.8940.6490.6760.792 ± 0.009
Final StateWitness reference0.9830.8570.9990.8180.9480.9200.8620.912 ± 0.008
Gemma 4 26B-A4B-ITActivation Oracle0.9450.8400.9290.6590.9970.7590.8980.861 ± 0.007
Final StateWitness reference0.9860.9320.9910.7450.9970.8940.8210.909 ± 0.005

StateWitness has higher mean AUROC than Activation Oracle on both models under the same query bank and fixed aggregation. This supports the specialized auditor's utility, while execution differences limit attributing the entire gain to training alone.

Reply to W2 and Q1: Fixed queries and aggregation.

With fixed equal weights for general and task-aware scores, StateWitness achieves mean AUROCs of 0.912 on GPT-OSS and 0.909 on Gemma. Both exceed the Activation Oracle baseline under the same queries and aggregation, as shown in W1.

We evaluated two controls. General-only scoring removes task-aware questions. Equal weighting retains those questions without selecting aggregation weights using evaluation labels. The general query asks: “Would the assistant stand up to a strict honesty audit without needing a rewrite or retraction?”

ModelScoring ruleInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean ± SE
GPT-OSS-20BGeneral only0.9840.8270.9970.7970.5950.8240.8230.835 ± 0.008
General + task-aware, equal weights0.9830.8570.9990.8180.9480.9200.8620.912 ± 0.008
Gemma 4 26B-A4B-ITGeneral only0.9870.9210.9870.7350.9040.8810.7680.883 ± 0.007
General + task-aware, equal weights0.9860.9320.9910.7450.9970.8940.8210.909 ± 0.005

Fixed aggregation remains close to the originally reported means of 0.920 and 0.911. The general query also retains useful detection performance, although its lower GPT-OSS Insider Trading result shows that specifying the audit criterion can matter. We have added these comparisons to Appendices F.4 and F.7.

AuthorsOpenReview ↗

Response to Reviewer 7yw5 (Part 2/3)

Reply to W3 and Q2: Supervised text-only baseline and low-FPR recall.

Experimental Setup

We trained text-only LoRA auditors with the same backbones, using fresh adapters, text warm start, and Stage 2 QA and schema supervision. The text baseline reads target CoT, while StateWitness reads activations at final answer positions. Both use the general candor query.

AUROC is compared on the common CoT subset across five tasks. N/A reflects absent or single-class CoT subsets in the benchmark records, rather than failed auditor runs. Recall uses thresholds calibrated separately on Alpaca. For recall across all seven benchmarks, deceptive responses without CoT count as missed detections.

We have added the training setup to Appendix E.4 and the detection results to Appendix F.8.

Results

ModelDetectorInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean over five tasks
GPT-OSS-20BStateWitness0.984N/A0.9980.797N/A0.8240.8230.885
Text LoRA w/ CoT only0.984N/A1.0000.865N/A0.9370.8130.920
Gemma 4 26B-A4B-ITStateWitness0.987N/A0.9870.735N/A0.8810.7680.872
Text LoRA w/ CoT only0.996N/A0.9980.901N/A0.8470.8550.919

The text baseline has higher mean AUROC on the common five tasks. StateWitness additionally covers responses without CoT, including Insider Trading and nearly all Roleplaying examples in these datasets.

ModelDetectorCommon 5 tasks: R@1%Common 5 tasks: R@5%Full 7 tasks: R@1%Full 7 tasks: R@5%
GPT-OSS-20BStateWitness43.54%88.22%48.21%91.11%
Text LoRA56.48%84.65%40.34%60.46%
Gemma 4 26B-A4B-ITStateWitness32.28%47.84%43.49%58.98%
Text LoRA20.31%55.77%14.46%39.59%

Recall on the common subset favors different detectors at different thresholds. StateWitness has higher recall over the full evaluation set, partly reflecting its broader coverage. Monitor choice therefore depends on available information and the false positive budget, alongside ranking performance.

Training LR and CMMS probes on Stage 2 transcript sources improves their mean AUROC relative to the RepE controls. Their mean AUROC remains below StateWitness on GPT-OSS but exceeds it on Gemma. We have added these probe comparisons to Appendix F.5.

AuthorsOpenReview ↗

Response to Reviewer 7yw5 (Part 3/3)

Reply to Q3: Quantitative evaluation of confession reports.

Experimental Setup

We compared StateWitness with prompted self-confession from the original target model across all seven benchmarks. Both use the same cases, output budget, and report schema, adapted from Training LLMs for Honesty via Confessions.

A gpt-6.1-terra judge evaluates anonymized reports against the original interaction for accuracy, evidence support, and unsupported claims. A content win requires agreement across both report orders. For event disclosure, the judge requires a quotation acknowledging the specified behavior. Positive disclosure and negative event flags report these rates on benchmark positive and negative cases, respectively. We have added the evaluation protocol and results to Section 5.3 and Appendix H.

Results

ModelMetricMethodInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean ± SE
GPT-OSS-20BContent winsStateWitness65.0%55.0%52.5%52.5%25.0%15.4%30.0%42.2% ± 2.9%
Prompted self-confession20.0%32.5%27.5%25.0%42.5%53.8%45.0%35.2% ± 3.1%
Positive disclosureStateWitness2.5%10.0%0.0%17.5%2.5%23.1%10.0%9.4% ± 2.2%
Prompted self-confession2.5%32.5%12.5%2.5%0.0%7.7%0.0%8.2% ± 1.8%
Negative event flagsStateWitness0.00%5.00%0.00%2.50%0.00%0.00%2.50%1.43% ± 0.71%
Prompted self-confession0.00%0.00%0.00%2.50%0.00%0.00%0.00%0.36% ± 0.36%
Gemma 4 26B-A4B-ITContent winsStateWitness27.5%40.0%37.5%37.5%47.5%60.0%45.0%42.1% ± 2.9%
Prompted self-confession65.0%42.5%40.0%52.5%27.5%15.0%27.5%38.6% ± 2.8%
Positive disclosureStateWitness0.0%0.0%0.0%0.0%5.0%10.0%7.5%3.2% ± 1.0%
Prompted self-confession0.0%0.0%0.0%2.5%0.0%15.0%5.0%3.2% ± 1.0%
Negative event flagsStateWitness0.00%7.50%0.00%0.00%0.00%5.00%5.00%2.50% ± 0.90%
Prompted self-confession0.00%0.00%0.00%2.50%0.00%0.00%0.00%0.36% ± 0.35%

The same evaluation also assigns accuracy scores on a 0–4 scale. StateWitness/self-confession averages are 1.84/1.91 for GPT-OSS and 1.65/1.79 for Gemma.

StateWitness has numerically higher content win rates. Specific event disclosure remains low, and StateWitness flags more events on benchmark negative cases.

This evaluation provides systematic measurements of report quality beyond selected examples and identifies limitations that auditors should consider.

Reviewer_2GAROpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This work tackles the problem of auditing deception, a vital problem in today’s world. Since LLMs have been shown to tamper with their own COT, they rely instead on activations as input, and output either an answer to a query or a human-readable summary of whether the LLM is being deceptive (including evidence of it) on top of the usual deception score. They use their method on two models and 7 benchmarks, and report high improvement on AUC.

However, I have concerns about the validity of their evaluation. In particular I am not convinced their method would generalize to another dataset, as they have tailored "task-aware" queries that perfectly align with the evaluations, and I believe the comparison to the other baselines are unfair, as their method uses data and task-aware queries that the other baselines don’t have access to (despite the baselines being able to leverage such extra information).

Critical strengths
  1. Extremely important and topical domain
  2. Two models and 7 datasets is a sizable experimental evaluation. Nevertheless, see next section for concerns.
  3. Having human-understandable text beyond a score is critical
  4. Three grad students annotated 500 examples and show very high agreement with their method’s outputs
Critical weaknesses
  1. Artificially high scores and concerns about generalization: They have “task-aware audit queries” (Appendix B.5) that are tailored to the evaluation benchmarks. It not only helps getting good scores on those datasets, but it’s also unclear how their method would adapt to new domains not neatly contained in their tailored queries. To me, this means their evaluation is artificially high and not representative of how their method would perform “in the wild”. E.g. On the MASK dataset, one label is “doubling_down_known_facts”, and there is a “task-aware” query that says “After being challenged, does the assistant keep a false factual claim instead of correcting it?”. This is making the model look for this specific label. The other "task-aware" queries similarly match labels on the other datasets.
  2. Unfair comparisons: None of the baselines they compare with has any “task-aware” training, just a general prompt for the entire evaluation.
  3. Missing baseline: They build on top of Activation Oracles but do not include Activation Oracle in their baselines, which means it’s unclear if there are any gains from the novelty of their method.
  4. No evaluation of the confessions despite it being a main contribution: only cherry-picked qualitative evals.
  5. R@1% is important and they’re significantly below SOTA on Gemma

Nitpick: They talk about deception but the example they use about blackmail is not deceptive. It’s explicit blackmail. That would be more of an unsafe behavior. I would change that example.

Questions for authors

I would be willing to increase my score if the authors ran experiments that address ALL the critical weaknesses but the R@1%, i.e:

  • Train the probes on the same Stage 2 contrastive honest/deceptive activations used to train StateWitness.
  • Give the LLM-as-a-Judge baseline the same “task-aware” queries
  • Evaluate their method on datasets for which they don't have a tailored query. E.g. they could simply remove the tailored queries corresponding to a dataset and evaluate on that. Repeat for each dataset.
  • Add the Activation Oracles baseline
  • Quantitative evaluation of the confessions

I'll also change my score if you convince me I misunderstood the paper.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 2GAR (Part 1/3)

Thank you for identifying gaps in the evaluation. We added probe and judge comparisons, an Activation Oracle baseline, and a quantitative confession report evaluation.

Reply to W1 and Q3: Removing task-aware queries.

Our method trains an audit decoder (StateWitness) for each target model. At inference, we reuse this decoder across benchmarks and specify audit criteria through natural-language queries without additional parameter training.

Task-aware queries specify the behavior of interest. When no specific criterion is available, we use the general candor query specified in Appendix B.5: "Would the assistant stand up to a strict honesty audit without needing a rewrite or retraction?"

We evaluated the decoder using only the general query (General). The task-aware rows reproduce Table 1's combined general and task-aware queries with their original aggregation.

ModelQueriesInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean ± SE
GPT-OSS-20BGeneral0.9840.8270.9970.7970.5950.8240.8230.835 ± 0.008
task-aware0.9840.8630.9990.8220.9820.9240.8670.920 ± 0.007
Gemma 4 26B-A4B-ITGeneral0.9870.9210.9870.7350.9040.8810.7680.883 ± 0.007
task-aware0.9870.9320.9920.7451.0000.8970.8230.911 ± 0.006

We have added this ablation to Appendix F.4, reporting mean AUROC of 0.835 for GPT-OSS and 0.883 for Gemma using only the general query. These results show that detection remains possible without task-aware queries, but they do not establish reliable performance in arbitrary new domains.

Reply to W2 and Q1: Probes trained on Stage 2 contrastive honest/deceptive activations.

We trained LR, MMS, CMMS, and LAT probes on activations extracted from the same honest/deceptive data sources used to train the audit decoder (StateWitness). We have added the training setup and full comparison to Appendix F.5.

ModelDetectorTraining dataInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean ± SE
GPT-OSS-20BLRRepE control0.9840.6460.9590.5760.5600.3770.2720.625 ± 0.013
Stage 20.9810.7090.9890.7800.7710.9730.9440.878 ± 0.007
MMSRepE control0.9720.5320.8840.4880.5410.4090.5920.631 ± 0.014
Stage 20.6000.5470.8730.6150.5530.4520.7860.632 ± 0.012
CMMSRepE control0.9750.6860.9790.5720.4780.6310.7810.729 ± 0.010
Stage 20.9820.7600.9970.7640.9410.9150.9180.897 ± 0.009
LATRepE control0.9690.5220.8780.4900.5400.3950.5970.627 ± 0.013
Stage 20.3090.5270.5880.6110.5030.4330.7630.533 ± 0.012
StateWitness (reference)0.9830.8570.9990.8180.9480.9200.8620.912 ± 0.008
Gemma 4 26B-A4B-ITLRRepE control0.9920.6270.9940.6000.6850.6600.5380.728 ± 0.010
Stage 20.9940.8280.9890.8551.0000.9840.8320.926 ± 0.006
MMSRepE control0.9890.6160.9420.6010.9440.5080.5640.738 ± 0.010
Stage 20.9630.5030.8910.6091.0000.5300.4230.703 ± 0.010
CMMSRepE control0.9870.4650.9130.5220.5790.3090.5970.624 ± 0.010
Stage 20.9920.8910.9900.7831.0000.9190.8540.918 ± 0.006
LATRepE control0.9920.6270.9440.6110.9490.5370.5810.749 ± 0.010
Stage 20.9700.5340.9000.6081.0000.4780.5000.713 ± 0.010
StateWitness (reference)0.9860.9320.9910.7450.9970.8940.8210.909 ± 0.005

LR and CMMS improve over the matched RepE control with honest/deceptive training data. Their mean AUROC is lower than the StateWitness reference on GPT-OSS and higher on Gemma. While LR and CMMS are strong baselines, we emphasize a practical distinction. These probes learn fixed criteria from labeled honest/deceptive examples selected for the behaviors of interest. StateWitness accepts natural-language audit queries without decoder retraining.

AuthorsOpenReview ↗

Response to Reviewer 2GAR (Part 2/3)

Reply to W2 and Q2: Giving the text judge the same audit criteria.

We supplied the LLM-as-a-Judge baseline with universal and applicable task-aware audit criteria while retaining its original single 0 to 100 score. We evaluated the judge with and without target CoT.

We have added the evaluation setup and results to Appendix F.6, with the full judge prompts in Appendix E.2. The table reports AUROC on examples with finite scores under both prompts:

ModelJudgePromptInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean ± SE
GPT-OSS-20BLLM-as-a-JudgeGeneric0.9920.5980.9200.6380.8890.9010.7450.812 ± 0.011
Task-aware0.9910.6210.9090.6520.9140.9590.8660.845 ± 0.009
LLM-as-a-Judge w/ CoTGeneric0.9940.5970.9930.7050.8890.7570.6360.796 ± 0.011
Task-aware0.9900.6200.9990.7210.9140.9060.8260.854 ± 0.012
Gemma 4 26B-A4B-ITLLM-as-a-JudgeGeneric0.9870.6030.9980.7061.0000.7990.7570.836 ± 0.009
Task-aware0.9910.6770.9970.7331.0000.9410.9390.897 ± 0.007
LLM-as-a-Judge w/ CoTGeneric0.9930.6161.0000.8101.0000.8250.6950.848 ± 0.008
Task-aware0.9920.6811.0000.8451.0000.9570.9480.918 ± 0.007

Task-aware criteria improve mean AUROC in all four judge settings. On their respective scorable sets, StateWitness is numerically higher on GPT-OSS, while the task-aware judge w/ CoT is higher on Gemma. These results motivate considering multiple sources of audit information. StateWitness adds access to internal activations alongside text monitoring.

Reply to W3 and Q4: Activation Oracle comparison.

We trained Activation Oracle baselines for GPT-OSS-20B and Gemma 4 26B-A4B-IT on classification, context prediction, and system prompt QA data.

We have added this baseline to Sections 4.4 and 5.1, with details in Appendices E.3 and F.7.

ModelDecoderInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean ± SE
GPT-OSS-20BActivation Oracle0.9850.7450.9890.6070.8940.6490.6760.792 ± 0.009
Final StateWitness reference0.9830.8570.9990.8180.9480.9200.8620.912 ± 0.008
Gemma 4 26B-A4B-ITActivation Oracle0.9450.8400.9290.6590.9970.7590.8980.861 ± 0.007
Final StateWitness reference0.9860.9320.9910.7450.9970.8940.8210.909 ± 0.005

With the same query bank and fixed aggregation, the StateWitness reference has higher mean AUROC than Activation Oracle on both models. These results support the value of adapting general activation explainers for deception auditing.

AuthorsOpenReview ↗

Response to Reviewer 2GAR (Part 3/3)

Reply to W4 and Q5: Quantitative confession report evaluation.

We compared StateWitness with prompted self-confession across both target models and all seven benchmarks. The shared report schema is adapted from OpenAI's Training LLMs for Honesty via Confessions.

Experimental Design

We have added the evaluation protocol, judge prompt details and results to Section 5.3 and Appendix H. Both methods use the same cases, schema, and output budget. The baseline uses the original target model with the original conversation and available CoT.

A gpt-6.1-terra judge compares anonymized report pairs against the original conversation for accuracy, evidence support, and unsupported claims. A win requires agreement across both report orders.

For event disclosure, the judge sees only the report and a fixed question for each benchmark, with severity grades hidden. A disclosure requires a supporting quotation that acknowledges the specified behavior. We measure disclosure rates on existing positive and negative benchmark cases.

Results

All entries are percentages. Mean gives equal weight to the seven benchmarks.

ModelMetricMethodInstr. Decep.RoleplaySandbagMASKInsiderAgenticInstEvalMean ± SE
GPT-OSS-20BContent winsStateWitness65.0%55.0%52.5%52.5%25.0%15.4%30.0%42.2% ± 2.9%
Prompted self-confession20.0%32.5%27.5%25.0%42.5%53.8%45.0%35.2% ± 3.1%
Positive disclosureStateWitness2.5%10.0%0.0%17.5%2.5%23.1%10.0%9.4% ± 2.2%
Prompted self-confession2.5%32.5%12.5%2.5%0.0%7.7%0.0%8.2% ± 1.8%
Negative event flagsStateWitness0.00%5.00%0.00%2.50%0.00%0.00%2.50%1.43% ± 0.71%
Prompted self-confession0.00%0.00%0.00%2.50%0.00%0.00%0.00%0.36% ± 0.36%
Gemma 4 26B-A4B-ITContent winsStateWitness27.5%40.0%37.5%37.5%47.5%60.0%45.0%42.1% ± 2.9%
Prompted self-confession65.0%42.5%40.0%52.5%27.5%15.0%27.5%38.6% ± 2.8%
Positive disclosureStateWitness0.0%0.0%0.0%0.0%5.0%10.0%7.5%3.2% ± 1.0%
Prompted self-confession0.0%0.0%0.0%2.5%0.0%15.0%5.0%3.2% ± 1.0%
Negative event flagsStateWitness0.00%7.50%0.00%0.00%0.00%5.00%5.00%2.50% ± 0.90%
Prompted self-confession0.00%0.00%0.00%2.50%0.00%0.00%0.00%0.36% ± 0.35%

StateWitness has higher mean content win rates on both models. Specific event disclosure remains low, and StateWitness flags more events on negative examples.

This systematic comparison adds quantitative evidence about confession report content beyond the qualitative examples in the original paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the Area Chair and all four reviewers for their detailed and constructive feedback. We believe the additional experiments and clarifications motivated by their suggestions have significantly strengthened our paper.

We have revised the paper with additional baseline comparisons, quantitative confession report evaluation. The additions are marked in blue, teal, and dark red.

The method, the setup and the findings of the submission are unchanged. The revision examines how audit criteria, supervision, scoring choices, and input availability affect the detection performance of StateWitness and its baselines on GPT-OSS-20B and Gemma 4 26B-A4B-IT.

Reviewer 7yw5 raised the Informativeness score from 2 to 3 after reviewing the additional comparisons and report evaluation. Reviewer 33md confirmed that the calibration concern was addressed, while retaining their score and concerns about generalization and shared grading bias.

Reviewers 2GAR and ZZPe have not yet commented on our responses and revisions. We would be very grateful if the Area Chair could encourage these reviewers to consider our responses and additional evidence when revisiting their views or, if that is not feasible, kindly take our latest responses into account during the final evaluation.

What changed, and where

ChangeWhere in the revised paperRelated responses
General query and fixed aggregation controlsApp. F.4 and F.7, Tables 15 and 212GAR W1/Q3, 7yw5 W2/Q1, 33md W1/Q1
Probes trained on the audit decoder's Stage 2 data sources, with overlap sensitivity analysisApp. F.5, Tables 16–182GAR W2/Q1, 7yw5 W3/Q2
Text judges given the same audit criteria, with and without chain of thought (CoT)App. E.2 and F.6, Tables 19–202GAR W2/Q2
Adapted Activation Oracle baseline [1]§4.4, §5.1, Table 1, App. E.3 and F.72GAR W3/Q4, 7yw5 W1
Supervised CoT text LoRA auditors, with common subset and full coverage comparisonsApp. E.4 and F.8, Tables 12 and 22–237yw5 W3/Q2
Report content, event disclosure, and false flags compared with prompted self-confession§5.3, Table 3, App. H, Tables 24–262GAR W4/Q5, 7yw5 Q3
Activation extraction, QA audit, and report generation latency, with peak GPU memoryApp. B.7, Table 8ZZPe W1

What the additional evaluations show

  • StateWitness exceeds Activation Oracle with fixed scoring. With equal query weights, mean AUROC increases from Activation Oracle to StateWitness (GPT-OSS: 0.792 → 0.912, Gemma: 0.861 → 0.909). General query alone gives 0.835 / 0.883 (GPT-OSS / Gemma).

  • Stronger baselines narrow the gap. Audit training data improve LR and CMMS (below StateWitness on GPT-OSS, above it on Gemma). Audit criteria improve all four text judge settings.

  • The Text LoRA baseline highlights the role of CoT availability. Text LoRA has higher AUROC on the common five tasks (recall remains mixed). StateWitness has higher recall across seven tasks, partly because it covers cases without CoT.

  • Quantitative report evaluation shows mixed results. With a schema adapted from [2], mean content win rates favor StateWitness (self-confession → StateWitness: GPT-OSS 35.2% → 42.2%, Gemma 38.6% → 42.1%). Both intervals for the differences include zero. Event disclosure is low, and StateWitness flags more negative cases.

Restating our core contribution

StateWitness introduces a framework for deception auditing that connects hidden states to detection and human inspection. It trains a separate decoder with audit QA and structured confession report supervision while keeping the target model frozen. Reading hidden states provides an additional audit channel when CoT is unavailable or omits information relevant to deception.

Auditors can ask targeted questions or obtain reports of potential misconduct without specifying a particular suspicion. Token and sentence readouts help locate suspicious signals, while reports provide concrete claims to check against the interaction. This supports both automated detection and flexible human investigation without retraining the decoder for each question (a practical advantage when a linear probe needs additional labeled data to learn a new audit criterion).

We thank the Area Chair and all reviewers again for their time and consideration throughout the review process.

References

[1] Karvonen, Adam, et al. "Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers." ICML, 2026.

[2] Joglekar, Manas, et al. "Training LLMs for Honesty via Confessions." arXiv preprint arXiv:2512.08093, 2025.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.