- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper investigates how emotion-related representations influence the behaviour of LLMs & LLM-based agents. The authors propose E-STEER, which combines VAD representations with SAEs to identify and manipulate emotion-related latent features. The framework is evaluated on reasoning, code generation, text generation, safety and agent-based tasks. The authors report nonlinear relationships between emotional steering and task performance, with different VAD configurations producing different behavioural outcomes. Additional experiments examine controllability, robustness across datasets and models and an application to jailbreak defence.
Critical strengths
-
The paper is easy to read and follow.
-
Investigating whether emotion-related representations can influence model behaviour beyond response style is an interesting topic. Instead of focusing, as typically is done, on emotional text generation, the paper considers objective reasoning, safety and multi-step agent tasks.
Critical weaknesses
The paper does not clearly establish what E-STEER offers beyond applying existing steering techniques to emotion-related features. As a result, the methodological novelty is quite limited for ICLR.
I also have several concerns about the proposed steering approach:
-
In Eq. (3), the steering direction is computed using the signed offset . However, Eq. (4) normalises this direction and then multiplies it by again. Assuming an affine SAE decoder, the signs would cancel for negative , meaning that positive and negative offsets of the same magnitude would produce steering in the same direction. The expression is also undefined when . This seems inconsistent with the positive and negative steering experiments presented in the paper. The authors should clarify how this is actually implemented. This is quite important, as it concerns the core of the proposed method.
-
The emotion-related latent features are identified through activation differences obtained using emotion-labelled prompts. However, I am not convinced that these differences necessarily correspond to emotion. They could also reflect changes in response style, instruction following, confidence, verbosity, etc. The random-latent experiment in Appendix A.2 is useful, but it is not sufficient enough to establish that the observed behavioural changes are specifically due to emotion rather than general activation steering.
-
Most experiments vary one VAD dimension at a time. There is little analysis of how the dimensions interact, whether they can actually be controlled independently, and/or whether steering one dimension also affects the others. Given that the unified and continuous representation of emotion is one of the main contributions, I would expect this to be investigated more thoroughly.
Another major concern is that the experimental results do not sufficiently support some of the conclusions drawn by the authors:
-
The statistical analysis is quite limited: confidence intervals and/or significance tests are not reported, which is particularly important given that some of the improvements are relatively small. Similarly, the claims about inverted-U relationships and optimal emotional states seem to be based mainly on visual inspection of the curves.
-
The authors often relate the observed performance trends to psychological mechanisms. However, these mechanisms are not directly investigated or measured in the experiments. The discussion should be more careful here, as observing similar performance patterns does not necessarily imply similar underlying mechanisms.
-
Several subjective and safety-related metrics rely on an LLM judge, but the evaluation protocol is not sufficiently described. Which model is used as the judge? What prompts are used and how reliable are the resulting evaluations? Also, in Appendix A.3, the VAD predictor used to evaluate controllability achieves correlations of only 0.55 and 0.51 for A and D, respectively. Given these values, I am not sure how reliable the near-perfect steering correlations reported in Table 4 are.
-
The comparison with existing methods is also limited. The controllability experiments mainly compare E-STEER with prompt-level control. Since the proposed approach builds on existing activation-steering techniques, I would also expect comparisons with relevant steering methods. Otherwise, it is difficult to determine whether E-STEER actually provides an improvement over existing approaches.
-
Finally, regarding the jailbreak defence experiments in Table 2, the improvements vary considerably across attacks and dominance values. The reported average ASR reduction of 29.1% appears to be calculated by choosing the best dominance setting for each attack separately. Were these values selected on a separate validation set? It would also be good to examine whether the improved defence comes at the cost of refusing legitimate requests. Without such an analysis, it is difficult to assess the practical benefit of the proposed approach.
Questions for authors
Please also have a look on the detailed critical weaknesses above:
-
Can the authors clarify the relationship between Eqs. (3) and (4), particularly how the sign of the steering offset is preserved? Was the direction computed using a fixed positive offset and subsequently scaled by signed values, or was the method implemented exactly as written?
-
How can the authors distinguish behavioural effects caused by emotional features from those caused by general representation perturbations? Have comparisons been conducted against non-emotional steering directions with matched intervention magnitudes?
-
To what extent can V, A and D be controlled independently? Does modifying one dimension affect the other two and what evidence supports effective steering of their joint configurations?
-
Can the authors provide evaluation sample sizes / confidence intervals / statistical tests for the main performance differences? Were the optimal steering configurations, particularly for jailbreak defence, selected and evaluated on separate data?
2 comments
I thank the authors for their rebuttal. I appreciate the additional experiments and clarifications.
My original concern about the steering direction is addressed (the revised formulation in Section 3.2). The additional experimental details and statistical analysis are also useful improvements.
However, I still have some concerns:
-
Statistical evidence: Table 6 shows that the improvements at valence +3 and +6 on LogiQA are not statistically significant compared with neutral. More importantly, none of the reported safety-risk probability comparisons on HarmBench is statistically significant. This raises questions about the strength of some of the main conclusions. The quadratic fits are also descriptive and do not establish the statistical significance of the claimed inverted-U relationships.
-
VAD control: The joint VA experiment is a useful addition, but it is limited to one task and does not establish independent control of the 3 dimensions. In addition, some of the selected neurons reported in Appendix C are shared across V and D.
-
Emotion-specific effects: Although the random-feature comparison is useful, I am still not fully convinced that the observed behavioural changes are specifically attributable to emotional representations rather than other correlated features (e.g., confidence, response style or general reasoning behaviour).
-
Contribution: Since the authors position the paper as an empirical study, I would still expect stronger evidence for the main empirical conclusions and clearer comparisons with existing steering approaches to establish the contribution.
The correction of the jailbreak results and the additional refusal-rate analysis are useful, but the latter is still limited in terms of benign-task coverage.
Overall, the revision addresses several of my initial concerns and improves the paper (and I appreciate that). However, the remaining concerns are still important to my assessment. Weighting everything, I will maintain my original rating.
We would thank the reviewer's construction advice, which helps us to refine the paper.
R1: Novelty and details of E-STEER (W0, W1, Q1).
We thank the reviewer for the question on novelty. We would like to clarify that this paper is positioned as an empirical study, to investigate how emotional signals affect the behavioral patterns of models. We do not claim to propose a new steering method; instead, we introduce E-STEER, a framework suited to our research goal.
Studying the effect of emotional signals requires a tool that can control them in a unified, continuous and fine-grained manner, which existing methods struggle to provide: prompting offers limited control over the signal; and existing steering typically relies on discrete labels, covering only a few emotions. Drawing on VAD theory, E-STEER maps discrete emotions into a continuous VAD space, extending the object of study from a few labels to the whole space, which allows us to vary intensity and dimensions systematically.
We acknowledge that the description of Eqs. (3)–(4) was insufficient. We have revised it; please see GR-1.
R2: Specificity of the emotional signal and interpretation (W2, W5, Q2).
We thank the reviewer for these comments. Regarding whether the behavioral changes are specific to emotion rather than general activation perturbation, please see GR-4, where we provide additional discussion of why the effects are unlikely to be explained by a simple generic signal.
Regarding the interpretation, we agree that our wording was too strong. We did not measure psychological mechanisms, and similar performance trends do not imply similar underlying mechanisms. In the revision, we have toned down the corresponding claims.
R3: Independence of VAD dimensions and measurement reliability (W3, W6, W7, Q3).
We thank the reviewer for the attention to the independence of the VAD dimensions. In our study, V, A and D are specified as independent input signals. We acknowledge that the dimensions may influence each other, and add this into limitation (Section 6). We added a joint valence × arousal experiment on LogiQA (V, A ∈ {−6, 0, +6}): the effect of valence depends on arousal and becomes negligible when arousal is high (new Appendix B.3).
Regarding the comparison with existing methods (new Appendix B.2), its purpose is only to rule out that E-STEER’s effect is merely prompt engineering. We admit that the VAD analyzer is only moderately accurate on the A and D dimensions, but its accuracy is on par with the original paper, which is sufficient to support the clear relative difference between the two methods. Existing steering methods mainly control discrete labels and adjust a semantic category direction rather than the VAD dimensions. Our metric is the correlation between steering intensity and the VAD values of the generated text, and it would be inappropriate to evaluate label-based steering with an analyzer designed for VAD, which is an unfair comparison. Our conclusions rely on E-STEER’s continuous, fine-grained control over VAD, not on it outperforming other baselines.
R4: Statistical analysis and evaluation protocol (W4, W6, Q4).
We acknowledge that the description of the statistical analysis and experimental setup was insufficient. We have now added this information; please see GR-2 and GR-3.
R5:Jailbreak implication and refusual rate (W8,Q4).
We thank the reviewer for the attention to the jailbreak impplication. Table 2 reports results at every fine-grained steering setting, and no training or tuning is involved; the issue lies in our reporting: the 29.1% figure averaged the best setting of each attack, which overstates the typical effect. We acknowledge this was a flaw in our reporting. In the revision, we average the relative ASR reductions across all three attacks and the five fine-grained dominance intensities (+4, +5, +6, +7, +8), yielding an average reduction of 14.0%.
Regarding refusal of legitimate requests, the refusal rate on LogiQA is 0 at all steering intensities along dominance, so benign requests are not refused. On jailbreak prompts, the refusal rate are higher than 85%, which reflects the intended effect of the defense. These results are added in Appendix D.