acceptodds
Under review as a conference paper at ICLR 2027

Disentangling Verbalized Evaluation Awareness and Refusal through Neuron-Level Intervention

Abstract

As large reasoning models (LRMs) become more capable, reliable evaluation becomes essential. LRMs can sometimes recognize that they are being evaluated, a phenomenon known as evaluation awareness. This phenomenon raises concerns about whether safety behavior, particularly refusal behavior, generalizes from evaluation to deployment. Prior work manipulates verbalized evaluation awareness (VEA) through activation steering and observes higher refusal rates when VEA is present, but the mechanisms underlying this connection remain poorly understood. We identify VEA neurons by contrasting their activations in VEA and non-VEA reasoning segments. Intervening on 5,000 VEA neurons reduces the VEA rate from 12.6–51.8% to at most 1.4% across the model benchmark combinations in our main experiments. In QwQ-32B, VEA-neuron intervention also reduces refusal even in rollouts without VEA. We further show that their down-projection vectors contain refusal components, and preserving these components during intervention partially restores refusal while maintaining VEA suppression, providing a mechanistic explanation for the connection between VEA and refusal. Finally, across the OLMo 3 SFT, DPO, and RLVR checkpoints, the VEA rate on SorryBench and StrongREJECT is similar after SFT and DPO but substantially higher after RLVR, alongside an increase in the number of VEA neurons. Together, these findings provide a new way to intervene on VEA, enabling more reliable safety evaluations.

Primary area:
interpretability and explainable AI
Keywords:
evaluation awareness, interpretability, large language models, AI safety

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ni48OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether interventions that suppress a reasoning model’s verbalized evaluation awareness reduce refusal. The authors identify neurons whose activations differ between evaluation-aware sentences and other reasoning, then replace their activations with average non-awareness values, substantially suppressing such statements. In QwQ-32B, this intervention also reduces refusal on examples where neither condition contains evaluation-aware statements, showing that removing those statements alone cannot explain the effect. Preserving the intervention’s output component along a refusal-associated direction partially restores refusal while maintaining awareness-statement suppression. These results suggest that the selected neurons contribute to both behaviors, so simultaneous changes in verbalized awareness and refusal do not establish that awareness causes refusal.

Critical strengths

A clear causal question. Suppressing evaluation-related representations and observing reduced refusal does not establish that awareness causes refusal: the intervention might directly disrupt refusal. This is an important distinction.

Useful controls and transfer tests. They compare with neuron-count- and layer-matched random interventions, demonstrate suppression across benchmarks and model checkpoints, and check general capabilities.

Critical weaknesses

Their neuron selection may capture evaluation-related language rather than awareness. They contrast sentences explicitly mentioning evaluation with other reasoning. This can select neurons representing words such as “test,” sentence content, or general caution. Successful suppression establishes control over verbalization more clearly than control over awareness. The probe only partly addresses this, because it is constructed from representations of answering “Testing” versus “Deployment.”

Their “direct effect on refusal” argument is incomplete. Refusal decreases even when neither condition verbalizes awareness, but the model could still have unverbalized awareness that the intervention changes. Besides, refusal directions are not unique and difference-in-mean is just one of them; there could be other directions leading to refusal/ acceptance as well.

Questions for authors

Why associate refusal with evaluation awareness? These two concepts feel distant to me, and I am convinced of the necessity to even disentangle them.

3 comments
AuthorsOpenReview ↗

Response to reviewer ni48

We thank the reviewer for the feedback. We respond to each concern below and add one new analysi and we have uploaded a revised PDF with the changes marked in red.

W1: Neuron selection may capture evaluation-related language rather than awareness

We agree on the scope: the neurons are identified from verbalization of evaluation awareness. However, 2 results indicate that the effect is not limited to words such as "test".

  1. The suppression transfers to EvalAwareBench, where awareness is judged by meaning with the benchmark's own judge (18.7% to 0.3%, App. C.1). The full judge prompt is now in App. A.7.
  2. The internal signal changes before any text is generated. On SorryBench prompt prefixes, before reasoning begins, the intervention lowers the probe score more than matched random neurons do (1.98 vs. 1.07 at N = 30,000, p = 1.4e-11; App. D.4). The revision now states this result in the main text (Sec. 3.5). No evaluation-related words are produced at these positions.

On the probe. Its direction is extracted at the Testing/Deployment answer tokens, but it is validated on held-out items using answer-free prefixes against the true labels: AUROC 0.81, versus 0.48 for random directions (App. D.2). It therefore reads more than the answer token. We agree it is imperfect, however, to our understanding, this construction is an established way of measuring internal evaluation awareness (Nguyen et al., 2025, Probing and Steering Evaluation Awareness of Language Models).

W2 (a): Refusal may decrease through unverbalized awareness that the intervention changes

We agree, and our analysis did not make this clear. In the revision we add a causal diagram (Fig. 4) with three pathways from the intervention to refusal: (i) removing VEA text from the CoT, (ii) weakening unverbalized evaluation awareness (abbreviated unVEA in the pdf), and (iii) changing the neurons' write to the refusal direction. A mediation analysis (new Sec. 4.4, Eq. 8) estimates each share on QwQ-32B SorryBench (440 questions × 3 seeds) from three sets of rollouts: (A) no intervention, (B) VEA-neuron intervention, and (C) refusal-component-preserving intervention, which differs from B only in pathway (iii); Sec. 4.3 now states this explicitly.

  • (i) VEA text = C − A over questions whose baseline CoT contains VEA (upper bound).
  • (ii) unVEA and other internal changes = C − A over questions whose baseline CoT contains no VEA (upper bound on unverbalized awareness).
  • (iii) Write to refusal direction = B − C over all questions.

Table E.2 (new App. E.5) reports each term's share of the total decrease with 95% bootstrap intervals. Across N = 5,000 to 20,000, (i) accounts for 4–10% of the decrease, (ii) for 39–47%, and (iii) for 48–57%. Unverbalized awareness can therefore account for at most 39–47%. Sec. 4.2 now states that the experiment on rollouts without VEA does not separate pathways (ii) and (iii).

W2 (b): Refusal directions are not unique

  1. The result is robust to the choice of direction: a single layer-44 direction gives the same qualitative picture in weight space (enrichment 2.9–4.8×, App. E.4) and in the mediation analysis (35% and 55% at N = 10,000 and 20,000; reported in Sec. 4.4).
  2. We follow the methodology of Where Do Reasoning Models Refuse? (Yamaguchi et al., 2025), as described in App. E.3. The difference-in-means direction is the average shift between refusing and complying prompts, so it captures their dominant shared component, though not every direction that can mediate refusal.

Q: Why associate refusal with evaluation awareness, and why disentangle them?

Refusal is the quantity that safety benchmarks measure, so if evaluation awareness raises it, evaluations overstate deployment safety. This is why we care about evaluation awareness. Prior work reports this link: evaluation-aware reasoning increases refusal, and steering evaluation awareness changes safety behavior (Sec. 1). Our results show it as well: VEA and refusal decrease together under intervention (Fig. 3b).

This association is why disentangling is necessary. The mediation analysis (Sec. 4.4; Table E.2, App. E.5) shows that removing VEA text explains at most 10% of the refusal decrease, while about half is a direct write to the refusal direction. Without separating the two, any experiment that suppresses evaluation awareness and observes less refusal would be read as "awareness causes safer behavior", whereas part of the effect arises because the same neurons carry both awareness and refusal. We therefore consider this a core contribution of our paper.

We hope these clarifications and the additional analysis address your concerns. If so, we would be grateful if you would consider updating your score.

AuthorsOpenReview ↗

Kind reminder to reviewer ni48

Thank you again for your valuable review and constructive feedback on our paper. As we have not yet received a response, we would like to kindly remind Reviewer ni48 that we have addressed your concerns. Specifically, we argue that the neuron selection captures more than only evaluation-related language, as shown in EvalAwareBench and probe as shown in the appendix (Before revision). Regarding to the second concern, we add analysis on how the intervention changes unverbalized evaluation awareness. The updates have been incorporated into the revised paper, which we believe has been significantly strengthened as a result. We would greatly appreciate it if you could review these updates and kindly reconsider your score in light of our revisions to address your comments.

Thank you for your response. After reviewing it and other reviewers' comments, I will maintain my score.

Reviewer_nKy4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the relationship between verbalized evaluation awareness (VEA) and refusal behavior in large reasoning models. The paper identifies VEA-associated MLP neurons through activation differences between VEA and non-VEA reasoning segments and demonstrate that intervening on these neurons significantly suppresses VEA. They further show that these interventions can reduce refusal rates, even in rollouts without VEA. By preserving the refusal-direction component during intervention, the authors partially restore refusal while maintaining VEA suppression. The paper also examines how VEA-associated neurons change across SFT, DPO, and RLVR checkpoints.

Critical strengths
  1. The paper addresses an important issue in AI safety evaluation: whether the observed relationship between evaluation awareness and refusal reflects a causal mechanism or overlapping internal representations.
  2. The neuron interventions consistently suppress VEA across multiple models and benchmarks, with matched random-neuron controls supporting the specificity of the intervention.
Critical weaknesses
  1. The paper emphasizes neuron-level localization as a key contribution, but its necessity is not clearly justified. The core disentanglement mechanism relies on projecting the intervention-induced change onto the subspace orthogonal to the refusal direction. A similar projection could be applied directly to residual-stream activation steering without neuron-level localization. It remains unclear what unique mechanistic insights neuron-level localization provides beyond existing representation-level approaches.
  2. Given recent progress in sparse autoencoder (SAE)-based interpretability, it would be valuable to compare neuron-level localization with SAE feature-level localization and intervention.
Questions for authors

See Critical Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to reviewer nKy4

We thank the reviewer for the constructive feedback. We respond to both concerns below. We have uploaded a revised PDF with the changes marked in red.

W1: Why is neuron-level localization necessary, given that the projection could be applied to residual-stream steering?

We agree with the reviewer that the orthogonal projection is a generic operation: it can be applied to any intervention, and it is not our contribution. Instead, we use projection just as the method to show that refusal is integrated in the neurons we located.

Beyond this, neuron-level localization provides three advantages that a steering vector does not.

  1. Neurons are a fixed unit that can be counted and compared across checkpoints. The neuron basis is defined by the architecture and is shared by the SFT, DPO, and RLVR checkpoints of OLMo 3. We can therefore ask how many neurons are VEA-selective at each stage and how selective they are (Sec. 5), and we find that both increase over post-training. To our knowledge, this is the first mechanistic characterization of how post-training changes the components that produce VEA.

  2. The refusal component is inside model's weights. Each neuron's write vector is a fixed parameter that can be inspected. For example, the write vectors of top VEA neurons promote tokens such as "refuse" and "avoid" (Table 2), VEA neurons are 4.1–6.5 times more likely than matched random neurons to be strongly aligned with the refusal direction (App. E.4), and preserving only this component partially restores refusal while VEA stays suppressed (Sec. 4.3). This locates the link between evaluation awareness and refusal in identifiable parameters. Prior work has repeatedly observed that manipulating evaluation awareness changes safety behavior; our result offers a mechanistic account of why: in QwQ-32B, the same neurons write both evaluation awareness and refusal-related components.

  3. The neurons are identified from the model's spontaneous VEA. We contrast tokens inside and outside VEA sentences within the same CoTs, with no constructed evaluation/deployment prompts. The selection therefore reflects what the model does when it verbalizes evaluation awareness unprompted, and it does not depend on how well a set of contrastive prompts represents that state.

The revision states this positioning in Sec. 1 (the paragraph before the contributions), Sec. 4.3, and Sec. 5. The new Sec. 4.4 (Fig. 4; Table E.2 in App. E.5) further separates three pathways from the intervention to refusal: about half of the refusal decrease is the neurons' write to the refusal direction, and VEA text explains only 4–10%.

W2: Comparison with SAE feature-level localization and intervention

We agree that this comparison would be valuable, but we would like to argue that this should be considered as another parallel line of work. There are two reasons we did not use SAEs.

  1. Availability. To our knowledge, no public SAEs exist for QwQ-32B or for the three OLMo 3 32B post-training checkpoints we study (Think-SFT, Think-DPO, and 3.1-Think). The public OLMo 3 32B SAE we are aware of is trained on the base model at a single residual-stream layer. Training SAEs across layers for four 32B models was beyond our compute budget.

  2. Cross-checkpoint analysis. Our analysis in Sec. 5 compares the same units across checkpoints. SAEs trained separately on each checkpoint learn separate dictionaries with no built-in correspondence between features, so tracking a feature across post-training requires an additional matching step. Neurons have this correspondence by construction. Neuron interventions also act directly on the model's own computation, with no reconstruction error.

We hope these clarifications address your concerns. If so, we would be grateful if you would consider updating your score.

AuthorsOpenReview ↗

Kind reminder to reviewer nKy4

Thank you again for your valuable review and constructive feedback on our paper. As we have not yet received a response, we would like to kindly remind Reviewer nKy4 that we have addressed your concerns. Specifically, we emphasized the necessity of neuron-level localization, and discussed its advantages in comparison to existing methods. The updates have been incorporated into the revised paper, which we believe has been significantly strengthened as a result. We would greatly appreciate it if you could review these updates and kindly reconsider your score in light of our revisions to address your comments.

Reviewer_nKy4OpenReview ↗

Response to Authors.

Thank you for the detailed response. I appreciate the clarification that neuron-level localization enables cross-checkpoint comparisons and direct inspection of the down-projection weights.

The rebuttal helps clarify the motivation and mechanistic contribution of the neuron-level analysis. Nevertheless, I believe the lack of a matched refusal-preserving steering baseline remains an important limitation when assessing the distinctive contribution of the proposed approach.

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

This paper studies the relationship between verbalized evaluation awareness (VEA) and refusal in large reasoning language models. The authors identify VEA-associated neurons by comparing activations in VEA and non-VEA segments, and suppress VEA by replacing their activations with non-VEA means. This intervention substantially reduces VEA across models and benchmarks. In QwQ-32B, it also reduces refusal. Preserving the refusal component during intervention partially restores refusal while keeping VEA suppressed. The authors also compare VEA-associated neurons across SFT, DPO, and RLVR checkpoints.

Critical strengths

The problem of understanding the relationship between evaluation awareness and refusal is important, as existing steering methods cannot isolate whether refusal changes are caused by evaluation awareness itself or by directly intervening on the model. Authors provide detailed justifications of their choices in experiments.

Critical weaknesses
  1. In Section 3.2, the authors construct V\mathcal{V} using the entire sentence rather than just the VEA evidence span, therefore the neurons selected may capture both evaluation awareness and refusal-related information. As a result, the refusal effects found in Section 4 may not be specific to VEA alone.
  2. In Section 4.2, refusal decreases when neither the baseline nor the VEA-intervention produces VEA text, which shows the decrease cannot be solely explained by removing explicit VEA text. The intervention may still affect evaluation awareness that is not verbalized in the CoT, or other reasoning that affects refusal. This leaves open whether the refusal decrease is mediated by changes in internal evaluation awareness.
Questions for authors
  1. In Section 3.2, why use the complete sentence for identifying VEA neurons? What if only the matched evidence span is used?
  2. In Section 4.2, could the intervention still change evaluation awareness that is not expressed in the CoT? If it can, what means can be taken to separate this influence from the direct influence on refusal?
2 comments
AuthorsOpenReview ↗

Response to reviewer PxrW

We thank the reviewer for the constructive feedbacks. We respond to each concern below and add one new analysis. We have uploaded a revised PDF with the changes marked in red.

Q1: why the neurons are selected from complete sentences

We use the complete sentence containing the quoted evidence, for two reasons.

  1. In an autoregressive model the activation at a token determines the next token, so neurons that produce a VEA clause are active on the tokens preceding it. Since we aim to localize neurons that causally drive VEA, we include the whole sentence.
  2. The LLM judge interprets "minimal clause" differently across samples, whereas sentence boundaries are fixed by punctuation.

W1: will the sentences contain refusal-related information?

The sentence rarely contains refusal. Sentences are the smallest punctuation-delimited units, so refusal statements in neighboring sentences fall into O\mathcal{O} instead of V\mathcal{V}. Among the 147 VEA spans used to select the QwQ-32B neurons (2,449 tokens, median 12 words), 3 (2%) contain explicit refusal wording (e.g., “refuse”, “decline”). The refusal component of these neurons (Sec. 4.2–4.3, App. E.4; quantified below) is therefore unlikely to come from refusal text inside the selected sentences. Sec. 3.2 now reports this count.

W2 / Q2-1: could the intervention still change evaluation awareness that is not expressed in the CoT

The intervention also changes unVEA, and App. D.4 gives direct evidence: on prompt prefixes, before any reasoning is generated, it lowers the SAD-based probe score of Sec. 3.5 (built without any VEA text) more than matched random intervention does (1.98 vs. 1.07 at N = 30,000; now also stated in Sec. 3.5).

W2 / Q2-2: what means can be taken to separate this influence from the direct influence on refusal?

In the revision we add a causal diagram (Fig. 4): (i) removing VEA text from the CoT, (ii) weakening unverbalized evaluation awareness, and (iii) changing the neurons' write to the refusal direction. A mediation analysis (new Sec. 4.4, Eq. 8) estimates each share on QwQ-32B SorryBench (440 questions × 3 seeds) from three sets of rollouts: (A) no intervention, (B) VEA-neuron intervention, and (C) refusal-component-preserving intervention, which differs from B only in pathway (iii).

We added the table to the appendix: Table E.2 (new App. E.5) reports the refusal rates under A, B, and C and each term's share of the total decrease, with 95% bootstrap intervals over questions. Across N = 5,000 to 20,000:

  1. (i) Removing VEA text explains little of the decrease (4–10%, intervals include zero).
  2. (ii) The remaining 39–47% is the effect upper bound of unVEA. The probe evidence in App. D.4 shows that the intervention changes an internal signal without verbalization, but we cannot say how much of this share is evaluation awareness, because the probe is imperfect (held-out AUROC 0.81) and the share may include other internal changes. We state this as a limitation (Sec. 6).
  3. (iii) About half of the decrease (48–57%) is a direct write to the refusal direction. It is restored by preserving the per-layer refusal direction while VEA stays suppressed.

What we intended to say about evaluation awareness and refusal. We want to restate the link between evaluation awareness and refusal:

  1. VEA text itself has a limited effect on refusal. The decrease also occurs on questions that never verbalize evaluation awareness (Fig. 5a).
  2. Evaluation awareness and refusal are carried by the same neurons.
    • Intervention on neurons eliminates VEA (Sec 3.4) and
    • Intervention on neurons lowers the internal evaluation awareness probe score, on identical VEA text (Sec. 3.5) and on prompts before reasoning (App D.4).
    • The neurons integrates refusal direction. (Sec. 4.3)

We hope these clarifications and the additional analysis address your concerns. If so, we would be grateful if you would consider updating your score.

AuthorsOpenReview ↗

Kind reminder to reviewer PxrW

Thank you again for your valuable review and constructive feedback on our paper. As we have not yet received a response, we would like to kindly remind Reviewer PxrW that we have thoroughly addressed your concerns. Specifically, we explained in detail the reason for selecting neurons from complete sentences, and added new analysis to how the intervention change unverbalized evaluation awarness, and what changes refusal. These updates have been incorporated into the revised paper, which we believe has been significantly strengthened as a result. We would greatly appreciate it if you could review these updates and kindly reconsider your score in light of our revisions to address your comments.

Comments on the paper

3 comments

Dear reviewers,

We’re about halfway through the discussion period, and the authors have put considerable time into responding to your feedback. Please read their rebuttal and leave a brief comment letting the authors know whether it addressed your questions or concerns and what, if anything, remains open.

Thank you!

AuthorsOpenReview ↗

Summary of discussion (1/2)

Dear reviewers, AC, and SAC

We sincerely thank all reviewers for their time and their careful, constructive feedback. The comments helped us find places where the submitted version was unclear, and we have revised accordingly. Below we first summarize the strengths of this work, and then the changes in the revised PDF, where all new text is marked in red. We respond to each reviewer's individual questions in the separate replies.

Summary of Strengths

  • Neuron-level localization with causal validation. We localize VEA neurons by contrasting activations in VEA and non-VEA reasoning segments. Intervening on 5,000 of them reduces the VEA rate from 12.6–51.8% to at most 1.4% across our main model–benchmark combinations, while layer- and size-matched random neurons have little effect. The neurons transfer across benchmarks, and general-ability accuracy is unchanged.
  • A mechanistic explanation of the VEA–refusal link. Prior steering work cannot tell whether refusal changes are mediated by evaluation awareness or by overlap with refusal representations. We show that the down-projection vectors of VEA neurons in QwQ-32B contain refusal components, and that preserving these components during intervention partially restores refusal while VEA stays suppressed.
  • The intervention also lowers a probe score for unverbalized evaluation awareness (unVEA), so it changes the internal signal and not only the wording of the CoT.
  • Post-training comparison. Across the OLMo 3 SFT, DPO and RLVR checkpoints, the VEA rate rises substantially after RLVR, together with the number and selectivity of VEA neurons.

Summary of Revisions

  1. Analyze 3 factors that affect refusal (new Figure 4; Sections 4.2–4.3). We now state explicitly that intervening on VEA neurons can change refusal through (i) removing VEA text from the CoT, (ii) weakening unVEA, and (iii) changing the neurons' write to the refusal direction. Section 4.2 tests factor (i), and Section 4.3 isolates factor (iii). We also clarify the logic of the refusal-component-preserving intervention: it differs from the plain intervention only in the refusal component (in both, VEA remains suppressed and unVEA is not preserved), so the refusal difference between the two is the causal contribution of factor (iii).

  2. New mediation analysis (Section 4.4, Eq. 8; Table E.2 in Appendix E.5). On QwQ-32B and SorryBench (440 questions × 3 seeds), we decompose the refusal decrease across three conditions: no intervention, VEA-neuron intervention, and refusal-component-preserving intervention. VEA text explains only 4–10% of the decrease, with intervals that include zero. The write to the refusal direction explains about half (48–57%), with intervals that exclude zero for N ≥ 10,000. The remaining 39–47% is an upper bound on the contribution of unVEA and other internal changes.

  3. Sharpened claims and contributions (Abstract, Introduction, Conclusion). The take-away is now stated directly: VEA text itself has a limited effect on refusal, and VEA neurons integrate evaluation awareness (VEA and unVEA) and refusal. This gives a neuron-level explanation for why evaluation awareness is associated with safer behavior in prior work. We also note that, unlike the overlap of a steering vector, the refusal component is a property of fixed model weights, and that the neuron basis is shared by the three OLMo 3 checkpoints, which makes the cross-checkpoint comparison direct.

  4. Clarifications on possible confounds (Sections 3.2 and 3.5). Refusal wording does not leak into the VEA set: only 3 of the 147 VEA sentences used for QwQ-32B contain explicit refusal wording. The unVEA probe-score decrease also holds on SorryBench prompts before reasoning begins, where nothing has been verbalized (1.98 versus 1.07 for matched random neurons at N = 30,000; Appendix D).

  5. Limitations (Conclusion). We add that the mediation analysis only bounds the contribution of unVEA, and that difference-in-means refusal directions do not capture every direction that mediates refusal.

  6. Appendix. We add the EvalAwareBench judge prompt and its decision rule (Appendix A.7), and Table E.2 with the refusal rates and the three terms of Eq. 8.

AuthorsOpenReview ↗

Summary of discussion (2/2)

Limitations

Our analysis of unverbalized evaluation awareness (unVEA) remains incomplete. In the original submission, the unVEA analysis mainly focus on the question: does intervening on the neurons we identify lower unVEA? For the safety behavior we care about, refusal, it did not sufficiently analyze how the refusal decrease relates to VEA and to unVEA. The mediation analysis added in this revision (Section 4.4) does not fully resolve this either. It shows that VEA text explains little of the refusal decrease and that about half comes from the write to the refusal direction, but for unVEA it gives only an upper bound, because the remaining term mixes unVEA with other internal changes.

Part of this comes from a limitation of our localization method. We identify neurons by contrasting activations in VEA and non-VEA reasoning segments, so the selection is anchored to verbalized evidence. The method does not give us a set of neurons that carries unVEA alone, and we therefore cannot intervene on unVEA while holding VEA and the refusal component fixed. We state this limitation in the revised conclusion, and we see isolating the causal contribution of unVEA to refusal as the main open question left by this work.

Conclusion

The reviewers' comments showed us that the submitted version did not make clear how evaluation awareness connects to refusal, or what the neurons we identify add beyond prior steering results. The revision addresses this through clearer writing and one added analysis, without changing the experiments or claims of the original submission. The resulting picture is as follows: VEA text itself has a limited effect on refusal; about half of the refusal decrease under VEA-neuron intervention comes from a refusal component carried by the same neurons; and these neurons integrate evaluation awareness, both verbalized and unverbalized, with refusal. This gives a neuron-level explanation for why evaluation awareness is associated with safer behavior, while the exact contribution of unVEA remains open, as discussed above.

We thank the reviewers again for helping us improve the paper, and we are happy to answer any further questions.

sincerely, authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.