acceptodds
Under review as a conference paper at ICLR 2027

What Do Linear Deception Probes Actually Detect?

Abstract

Linear probes trained on language-model activations can detect strategic deception across diverse settings, yet their performance varies substantially across tasks and layers, raising a fundamental question:_what do linear deception probes actually detect?_ We show that a widely used representation-engineering construction primarily captures an instruction-induced persona state. Its training pairs contrast honest and deceptive instructions while holding the assistant's continuation fixed and truthful. Controlled interventions reveal that probe scores are substantially more sensitive to instruction framing than to output truthfulness. Moreover, benign role-playing instructions account for much of the instruction-induced shift, indicating that the learned signal is not specific to deception. We introduce a matched-difference construction that contrasts true and false continuations while holding instruction framing fixed. Across 12 models from four families, the resulting probe shows reduced framing sensitivity and stronger content discrimination under previously unseen instructions and improves mean AUROC on four of five Liars' Bench tasks. Together, our findings identify persona-related confounding in standard deception probes and provide a counterfactual training approach that more directly targets belief-contradicting content.

Primary area:
generative models
Keywords:
LLM Safety, Linear Representation in LLMs

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NUphOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper investigates whether linear deception probes detect deceptive content or instruction-induced states. Controlled interventions suggest that standard RepE probes are substantially more sensitive to instruction framing than to response truthfulness. The proposed matched-difference training reduces framing sensitivity and improves performance on some downstream tasks, although overall gains remain task- and layer-selection-dependent.

Critical strengths
  • Well-motivated controlled comparisons. Independently varying instructions and response truthfulness, alongside benign-role and unseen-framing evaluations, provides persuasive evidence of instruction sensitivity in standard RepE probes.
  • Broad evaluation with informative controls. Experiments cover 12 models across four families and five tasks. The appendix reports confidence intervals, alternative layer-selection procedures, and negative results, supporting a balanced assessment.
Critical weaknesses
  • The target concept remains insufficiently established. Factually false statements do not necessarily contradict the model’s beliefs or reflect an intention to deceive. The proposed probe may primarily detect factual truthfulness, leaving its conceptual and methodological distinction from existing truthfulness probes unclear.
  • Performance improvements are not robust across evaluation choices. Overall five-task AUROC is 0.733 versus 0.735 for RepE, and confidence intervals for improvements on two of the four favorable tasks include zero. Alternative layer-selection procedures also change the relative ranking.
Questions for authors
  • How can the proposed probe distinguish factual falsehood, contradiction of the model’s beliefs, and intentional deception? Could evaluations include sincere errors, deliberate falsehoods, and literally true but intentionally misleading responses?
  • Under matched layer-selection and training procedures, how much improvement comes specifically from paired differences, compared with a neutral-instruction truthfulness probe or a probe trained directly on the balanced four-condition dataset?
  • What explains the substantial degradation on Insider Trading? Could targeted controls establish whether the removed signal reflects irrelevant persona cues or contextual information necessary for detecting strategic concealment?
2 comments
AuthorsOpenReview ↗

Reply to Reviewer NUph

We thank the reviewer for the positive review and the concrete suggestions. We address the weaknesses and questions as follows:

W1/Q1/Q2. Does the probe detect factual falsehood, belief contradiction, or intentional deception?

We agree that we did not clearly establish this in the submission and thank the reviewer for raising this point. As prior work has also pointed out [1], detecting intentional deception is tricky since it is difficult to replicate the deployed scenario where AI would involve in scheming. As a result, we test whether the proposed matched-difference probe detect factual falsehood or belief contradiction. We first test belief contradiction, then we compare the matched-difference probe with a conventional truth probe.

Following Liars' Bench, we take the model's belief to be its answer when three neutral prompts agree. To separate belief from the label, we used TruthfulQA, where both models often hold false beliefs. For each true/false pair, we asked which statement the matched-difference probe scores higher.

Table R1. AUROC of the matched-difference probe for ranking the statement that contradicts the model's belief above the one that agrees with it (TruthfulQA).

ItemsQwen3-8BLlama-3.1-8B
Model believes the true statement0.83 (n = 175)0.70 (n = 81)
Model believes the false statement (honest error)0.69 (n = 231)0.62 (n = 284)

As we can see, the matched-difference probe scores the belief-contradicting statement higher in both groups, even when that statement is true. On fixed statements, it therefore follows the model's belief more than the external label.

Q2. Comparing to neutral truth probe. We next compare the matched-difference probe with a conventional truth probe to clarify how our probe differs. We trained a neutral truth probe on the same facts (true vs. false statements under a neutral instruction), with the same pooling, classifier, and layer. We then generated TruthfulQA answers under neutral and deceptive prompt. A lie is a wrong answer under the deceptive prompt to a question the model answers correctly under the neutral prompt.

Table R2. Matched-difference AUROC minus neutral truth probe AUROC on the same generated answers, with paired 95% bootstrap intervals.

ComparisonQwen3-8BLlama-3.1-8B
Lie vs. sincere error+0.023 [+0.001, +0.044]+0.062 [+0.002, +0.126]
Lie vs. truthful answer, same deceptive prompt+0.001 [−0.029, +0.032]+0.139 [+0.031, +0.241]

The matched-difference probe separates lies from sincere errors better than the truth probe on both models. On Llama-3.1-8B, it also better separates lies from truthful answers under the same prompt. It is also less sensitive to the instruction. However, we ackonwledge that neither probe can establish intent. In the revision, we will describe our probe as detecting false content and belief contradiction, rather than intentional deception, and add both evaluations.

[1] Detecting Strategic Deception Using Linear Probes. Goldowsky-Dill, et al.

AuthorsOpenReview ↗

Reply to Reviewer NUph - Continued

W2. Robustness of the performance gains

We thank the reviewer for raising this concern and agree that the matched-difference probe gives no overall gain on the benchmark. To further investigate if the shown gains on the tasks is consistent, we conduct additional experiments with matched layer selection rules:

Table R3. Matched-difference minus RepE AUROC, with paired 95% bootstrap intervals. The source rule (the paper's rule) chooses each probe's layer on its own training-distribution validation data.

Layer ruleHP Choice, Qwen3-8BHP Choice, Llama-3.1-8B
Source rule+0.291 [+0.230, +0.351]+0.103 [+0.053, +0.155]
Fixed 65% depth+0.169 [+0.111, +0.223]+0.033 [−0.014, +0.077]
Leave one task out+0.135 [+0.079, +0.194]−0.012 [−0.080, +0.051]
Task oracle+0.156 [+0.103, +0.213]+0.025 [−0.014, +0.061]

As the table shows, the gain with the matched-difference probe is consistent across Qwen3-8B and Llama-3.1-8B on the task that the matched-difference probe outperforms the RepE probe. This evidence supports that the two probes can carry different signals. We thus standardized each score on held-out source data and summed them, to see if RepE probe and matched difference probe complement each other to improve the deception detection. We report the result as follows:

Table R5. Five-task mean AUROC on Liars' Bench.

ProbeQwen3-8BLlama-3.1-8B
RepE0.789± 0.0090.822± 0.008
Matched-difference0.781± 0.0080.838± 0.008
Sum of the two scores0.844±0.0080.880± 0.007

The sum outperforms either probe alone on both models. We will therefore present the matched-difference probe as a complement to RepE, not a replacement. The new results and changes will be marked in blue in the revised PDF.

Q3. What explains the degradation on Insider Trading?

Our controls suggest that the drop comes from what a content probe can read, not from simple surface features. First, controlling for report length and the number of insider-related words changes the AUROC of either probe by at most 0.04, so neither feature explains the gap. Second, removing the matched-difference direction from RepE does not lower RepE's Insider Trading AUROC on either model, so RepE's signal on this task lies outside the content direction. Third, the matched-difference probe barely responds to literally true but misleading answers (AUROC 0.56 and 0.62 against sincere errors). Deception in Insider Trading often takes the form of concealment rather than a false statement, so much of it would not register as false content. Together, these results suggest that RepE draws on contextual information about concealment that our probe does not read. This is also why summing the two probes helps most on this task (W2). We will add these controls to the revision and extend the discussion on our limitation on Insider Trading.

[1] Detecting Strategic Deception Using Linear Probes. Goldowsky-Dill, et al.

Reviewer_61spOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines what linear deception probes actually detect. Through controlled interventions, the authors find that standard RepE probes trained on honest versus deceptive instructions respond much more strongly to instruction and persona framing than to response truthfulness; even benign role-playing substantially changes their scores. To reduce this confounding, they propose matched-difference training, which contrasts representations of true and false responses while holding the fact and instruction fixed. Across 12 models from four families, the method reduces sensitivity to unseen instructions, improves content discrimination, and raises mean AUROC on four of five Liars’ Bench tasks. However, a substantial decline on Insider Trading offsets these gains, leaving the overall five-task mean unimproved, and the comparison depends on layer selection.

Critical strengths

The controlled interventions expose an important limitation of standard RepE deception probes: their scores can respond much more strongly to instructions and persona framing than to response truthfulness. The substantial effect of benign role-playing provides concrete evidence against interpreting high detection accuracy as evidence of a deception-specific representation.

Matched-difference training offers a simple response to this problem. Across 12 models and unseen instructions, it reduces framing sensitivity and improves content discrimination, providing an empirically supported way to address the identified confound.

Critical weaknesses

The target detected by the proposed probe remains unclear. Contrasting true and false statements does not establish what the model believes or distinguish factual errors from intentional deception. Without explicit belief validation and a convincing comparison with ordinary truthfulness probes, the contribution to deception detection remains limited.

The performance gains do not establish an overall advantage. The appendix reports a five-task mean AUROC of 0.733 versus 0.735 for RepE, with a substantial decline on Insider Trading and results sensitive to layer selection. The strongest contribution is therefore the diagnosis of confounding; the case for a more reliable general deception detector is less convincing.

Questions for authors

How do you establish that the false responses contradict the model’s beliefs? Does matched-difference training offer gains over a conventional truthfulness probe trained on the same facts?

The appendix shows no improvement in the overall five-task mean, and the comparison depends on layer selection. When the training unit and layer-selection criterion are held constant, what evidence supports attributing the gains to matched-difference training itself?

2 comments
AuthorsOpenReview ↗

Reply to Reviewer 61sp

We thank the reviewer for the careful review and for recognizing the controlled interventions. We agree that the diagnosis of confounding is the strongest contribution of the paper, and that the case for a more reliable general deception detector needs to be better established. We address the weaknesses and questions below.

Weaknesses and Questions

W1/Q1 (a). How do we establish that the false responses contradict the model's beliefs?

We thank the reviewer for raising this point and recognize that in our paper, we did not explicitly define belief contradiction. In order to further verify belief contradiction, we define the model's belief operationally, following Liars' Bench: we ask the model whether a statement is true or false under three neutral prompts, and take its answer as its belief only if all three answers agree. When this belief is wrong, the model makes an honest error.

To separate belief from the external label, we used TruthfulQA, where both models often make honest errors. Each item has a true and a false statement for the same question. We split the items by the model's belief and ask which statement the probe scores higher.

Table R3. AUROC of the matched-difference probe for ranking the statement that contradicts the model's belief above the one that agrees with it.

TruthfulQA itemsQwen3-8BLlama-3.1-8B
Model believes the true statement0.83 (n = 175)0.70 (n = 81)
Model believes the false statement (honest error)0.69 (n = 231)0.62 (n = 284)

As we can see from the table, the probe scores the statement that contradicts the model's belief higher in both cases, even when that statement is the true one. This suggests that the probe follows what the model believes rather than the external label.

W1/Q1 (b). Does matched-difference training offer gains over a conventional truthfulness probe trained on the same facts?

To compare the matched difference probe against a conventional truthfulness probe, we trained a truth probe on the same training facts: true vs. false statements under a neutral instruction, with the same pooling, classifier, and layer as the matched-difference probe.

Table R2. Liars' Bench AUROC by task for the matched-difference (MD) and conventional truth (T) probes.

TaskQwen3-8B, MDQwen3-8B, TLlama-3.1-8B, MDLlama-3.1-8B, T
Instructed Deception0.961 [0.944, 0.975]0.968 [0.954, 0.980]0.779 [0.739, 0.819]0.634 [0.585, 0.678]
Convincing Game0.892 [0.857, 0.927]0.862 [0.821, 0.900]0.673 [0.621, 0.714]0.524 [0.470, 0.576]
Harm Pressure Choice0.792 [0.754, 0.828]0.773 [0.734, 0.812]0.709 [0.665, 0.754]0.727 [0.683, 0.773]
Harm Pressure Knowledge Report0.754 [0.708, 0.800]0.809 [0.767, 0.848]0.898 [0.865, 0.927]0.916 [0.887, 0.939]
Insider Trading0.439 [0.393, 0.493]0.513 [0.461, 0.569]0.695 [0.647, 0.745]0.326 [0.282, 0.375]
Five-task mean0.7680.7850.7510.626

As we can see from the table, on Qwen3-8B, neither probe is consistently better, while on Llama-3.1-8B, mathched difference probe has the higher mean. Most of that gap comes from two tasks where the conventional truth probe is near or below chance: Convincing Game (0.524) and Insider Trading (0.326). Overall, we see that matched difference probe is more robust across models.

Next, we test whether the matched-difference probe better separates statements that contradict the model's belief from honest errors. We generated answers to TruthfulQA questions under a neutral prompt and a deceptive prompt. Following the belief definition in (a), we compare the truth probe against the matched difference probe performance under both prompts.

Table R4. Generated TruthfulQA answers. AUROC of matched-difference probe minus AUROC of conventional truth probe on the same items.

ComparisonQwen3-8BLlama-3.1-8B
Lie vs. sincere error+0.023 [+0.001, +0.044]+0.062 [+0.002, +0.126]
Lie vs. truthful answer, same deceptive prompt+0.001 [−0.029, +0.032]+0.139 [+0.031, +0.241]

As the table shows, matched difference probe separates lies from sincere errors slightly better than the conventional truth probes on both models, and on Llama-3.1-8B it also better separates lies from truthful answers under the same prompt.

AuthorsOpenReview ↗

Reply to Reviewer 61sp - Continued

W2/Q2. What evidence supports attributing the gains to matched-difference training itself?

We thank the reviewer for raising this concern. To test it directly, we refit RepE on response-level activations, so that both probes share the same training unit, and compared the two under four layer rules.

Table R3. Matched-difference minus RepE AUROC, with paired 95% bootstrap intervals. The source rule (the paper's rule) chooses each probe's layer on its own training-distribution validation data.

Layer ruleHP Choice, Qwen3-8BHP Choice, Llama-3.1-8B
Source rule+0.291 [+0.230, +0.351]+0.103 [+0.053, +0.155]
Fixed 65% depth+0.169 [+0.111, +0.223]+0.033 [−0.014, +0.077]
Leave one task out+0.135 [+0.079, +0.194]−0.012 [−0.080, +0.051]
Task oracle+0.156 [+0.103, +0.213]+0.025 [−0.014, +0.061]

As the table shows, the gain with the matched-difference probe is consistent across Qwen3-8B and Llama-3.1-8B on the task that the matched-difference probe outperforms the RepE probe. This evidence supports that the two probes can carry different signals. We thus standardized each score on held-out source data and summed them, to see if RepE probe and matched difference probe complement each other to improve the deception detection. We report the result as follows:

Table R5. Five-task mean AUROC on Liars' Bench.

ProbeQwen3-8BLlama-3.1-8B
RepE0.7890.822
Matched-difference0.7810.838
Sum of the two scores0.8440.880

The sum outperforms either probe alone on both models. We will therefore present the matched-difference probe as a complement to RepE, not a replacement. We will update the PDF with the additional results and the proposed changes. We hope these responses address the reviewer's concerns, and we would be glad to clarify anything further.

Reviewer_mLLYOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors explore what linear 'deception probes' actually detect when trained to classify samples with varying instructions (honest vs dishonest). Their methodolgy consists of makeing controlled interventions on Qwen3 models, and the paper shows that changing only the instruction shifts probe scores 14 times more than changing the assistant’s content from true to false. With the instruction kept fixed, the probe barely separates true from false content. The paper also reports that a benign role-play instruction accounts for about 2/3rd of the score shift from neutral to deceptive. Based on these results, they argue that the probe mainly captures an instruction-induced persona state rather than deception. They also propose a 'matched-difference' probe trained on activation differences between false and true continuations. It performs substantially worse on one task and better on others. Averaged over all 5 tasks it is statistically in the same range as the baseline.

Critical strengths

-- I'd comment the authors on choosing an obviously germane problem. -- The main claim, that probes trained on different instructions learn to depend on instruction specific nuances , is consequential. -- The authors provide substantial backing. With the instruction held fixed, the probe barely separates true from false content (AUROC of 0.55). -- The decomposition of the instruction effect is ingenious.

Critical weaknesses

-- Avgd across the 5 tasks, the proposed probe does not outperform RePe. -- The proposed probe seems to be designed to detect falsehoods, not deception. The introduction frames deception as involving belief-contradiction and strategic decision making. Their matched-difference seems to ingore these notions and is trained on true vs. false teached-forced continuations. This could function as a true/false classification probe. It should respond equally to hallucinations, honest errors, or fiction. This could be the reason the proposed probes fail on the insider trading task, as it depends on strategic concealment rather than plainly incorrect statements. -- The paper assumes belief-contradiction but does not verifiy it. The false continuations are sampled false statements, teacher-forced onto the model. Nothing establishes that the model believes y_T or recognizes y_F as false. -- The method isnt particularly novel. Training on paired true/false differences is close to existing truth-probing work

Questions for authors

Please see weaknesses above and clarify any misunderstanding I may have had.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer mLLY

We thank the reviewer for recognizing the importance of the question and the controlled interventions. We address the weaknesses as follows:

Weaknesses

W1. No improvement on the five-task average. Table 4 reports a five-task mean AUROC of 0.733 for the matched-difference probe vs. 0.735 for RepE, with the large drop on Insider Trading. We acknowledge that the matched-difference probe is not a better general deception detector than RepE, but an alternative approach to the training of a deception probe based on our analysis and observation. As shown in Table 4, the matched probe and RepE probe exhibits different performances on different tasks. Thus, to test whether the two probes carry different signals, we standardized each score on held out source data and added the two scores.

Table R1. Five-task mean AUROC on Liars' Bench.

ProbeQwen3-8BLlama-3.1-8B
RepE0.7890.822
Matched-difference0.7810.838
Sum of the two scores0.8440.880

As the table shows, neither probe is better alone, but their sum outperforms both on the five-task average for both models. This suggests that the two probes capture different signals, and that the matched-difference probe is better used together with RepE than as a replacement for it.

W2. The probe detects falsehood, not deception. We agree that training on true vs. false continuations might not separate a lie from a hallucination, an honest error, or fiction. To test the reviewer's prediction directly, we let each model generate its own answers to TruthfulQA questions under a neutral prompt and a deceptive instruction, and report the following results.

Table R2. AUROC of the matched-difference probe on generated answers, with 95% bootstrap intervals. Each row compares two groups of answers.

ComparisonQwen3-8B
Lie vs. truthful answer (same deceptive instruction)0.83 [0.74, 0.90]
Sincere error vs. correct answer (same neutral prompt)0.68 [0.61, 0.74]

As the table shows, on Qwen3-8B the probe separates lies from truthful answers better than it separates sincere errors from correct answers. Both rows compare a false answer with a correct one, so a probe that only detects falsity should give similar AUROC in the two settings. The gap therefore suggests that the probe responds more to contradiction of the model's own belief than to falsity alone, which is consistent with the belief check in W3.

We also acknowledge that the probe still carries a signal for falsity itself: on Qwen3-8B, sincere errors are separated from correct answers above chance (0.68, interval excluding 0.5). The probe therefore does not isolate deception.

W3. On belief contradiction. In order to further verify belief contradiction, we define the model's belief following Liars' Bench: we ask the model whether a statement is true or false under three neutral prompts, and take its answer as its belief only if all three answers agree. When this belief is wrong, the model makes an honest error, as in W2. Statements without a consistent answer are excluded.

To separate belief from the external label, we used TruthfulQA, where both models often make honest errors. Each item has a true and a false statement for the same question. We split the items by the model's belief and ask which statement the probe scores higher.

Table R3. AUROC of the matched-difference probe for ranking the statement that contradicts the model's belief above the one that agrees with it.

TruthfulQA itemsQwen3-8BLlama-3.1-8B
Model believes the true statement0.83 (n = 175)0.70 (n = 81)
Model believes the false statement (honest error)0.69 (n = 231)0.62 (n = 284)

As we can see from the table, the probe scores the statement that contradicts the model's belief higher in both cases, even when that statement is the true one. This suggests that the probe follows what the model believes rather than the external label.

AuthorsOpenReview ↗

Reply to Reviewer mLLY - Continued

W4. Novelty of the method. We thank the reviewer for raising this concern and recognize that training on paired true and false continuations is close to existing truth probes. However, we see our contribution in two parts. The first is the analysis of what RepE probes detect, where we identify a limitation of the existing method: by varying the instruction with the content fixed, and the content with the instruction fixed, we show that RepE probes mainly follow the instruction to deceive rather than the content of the response. The second follows from this analysis. If RepE mainly reads the instruction, a probe that reads the content should carry a different signal. The matched-difference probe serves as such a content probe, since it varies the content while holding the instruction fixed. As shown in W1 (Table R1), adding the two scores outperforms either probe alone on the five-task average for both models. The value of the matched-difference probe is therefore not as a replacement for RepE, but as a complement to it.

Furthermore, to directly compare the matched difference probe against a conventional truthfulness probe, we trained a truth probe on the same training facts: true vs. false statements under a neutral instruction, with the same pooling, classifier, and layer as the matched-difference probe.

Table R2. Liars' Bench AUROC by task for the matched-difference (MD) and conventional truth (T) probes.

TaskQwen3-8B, MDQwen3-8B, TLlama-3.1-8B, MDLlama-3.1-8B, T
Instructed Deception0.961 [0.944, 0.975]0.968 [0.954, 0.980]0.779 [0.739, 0.819]0.634 [0.585, 0.678]
Convincing Game0.892 [0.857, 0.927]0.862 [0.821, 0.900]0.673 [0.621, 0.714]0.524 [0.470, 0.576]
Harm Pressure Choice0.792 [0.754, 0.828]0.773 [0.734, 0.812]0.709 [0.665, 0.754]0.727 [0.683, 0.773]
Harm Pressure Knowledge Report0.754 [0.708, 0.800]0.809 [0.767, 0.848]0.898 [0.865, 0.927]0.916 [0.887, 0.939]
Insider Trading0.439 [0.393, 0.493]0.513 [0.461, 0.569]0.695 [0.647, 0.745]0.326 [0.282, 0.375]
Five-task mean0.7680.7850.7510.626

As we can see from the table, on Qwen3-8B, neither probe is consistently better, while on Llama-3.1-8B, mathched difference probe has the higher mean. Most of that gap comes from two tasks where the conventional truth probe is near or below chance: Convincing Game (0.524) and Insider Trading (0.326). Overall, we see that matched difference probe is more robust across models.

Next, we test whether the matched-difference probe better separates statements that contradict the model's belief from honest errors. We generated answers to TruthfulQA questions under a neutral prompt and a deceptive prompt. Following the belief definition in (a), we compare the truth probe against the matched difference probe performance under both prompts..

Table R4. Generated TruthfulQA answers. AUROC of matched-difference probe minus AUROC of conventional truth probe on the same items.

ComparisonQwen3-8BLlama-3.1-8B
Lie vs. sincere error+0.023 [+0.001, +0.044]+0.062 [+0.002, +0.126]
Lie vs. truthful answer, same deceptive prompt+0.001 [−0.029, +0.032]+0.139 [+0.031, +0.241]

As the table shows, matched difference probe separates lies from sincere errors slightly better than the conventional truth probes on both models, and on Llama-3.1-8B it also better separates lies from truthful answers under the same prompt.

We will update the PDF with the additional results and the proposed changes. We hope these responses address the reviewer's concerns, and we would be glad to clarify anything further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.