acceptodds
Under review as a conference paper at ICLR 2027

From Representation to Behavioral Control: An Empirical Study of Fine-Grained and Unified Emotional Signals in LLMs

Abstract

Emotions are critical for human cognition and behavior, motivating the study of similar emotional signals in LLMs. Existing work is limited by narrow evaluation scopes focusing on subjective generation, and discrete emotion representations that overlook continuous variations and relations among emotions. To address these limitations, we conduct a systematic empirical study on the effects of emotional signals, supported by **E-STEER**, an interpretable fine-grained emotion steering framework designed as a research instrument. E-STEER models emotions in VAD space, providing a continuous and unified representation. Building on this representation, E-STEER identifies emotion-related latent features and enables quantitative intervention in hidden states. With the framework, we examine the impact of emotion on diverse tasks. The results reveal non-monotonic emotion-behavior relations consistent with established psychological theories, demonstrating how emotional signals systematically shape model behaviors. We further demonstrate the practical value of this signal as a fine-grained behavioral control factor for LLM applications, with jailbreak defense as a representative case.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Emotion-aware Modeling, Large Language Model, LLM-based Agent, Representation-Level Steering, Mechanistic Interpretability, Sparse Autoencoder

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_aa9COpenReview ↗
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 δi\delta_i. However, Eq. (4) normalises this direction and then multiplies it by δi\delta_i again. Assuming an affine SAE decoder, the signs would cancel for negative δi\delta_i, meaning that positive and negative offsets of the same magnitude would produce steering in the same direction. The expression is also undefined when δi=0\delta_i = 0. 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

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.

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.

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

The paper proposes E-STEER, a framework for studying how emotion-associated representations influence LLM and agent behavior. It identifies sparse autoencoder features associated with valence, arousal, and dominance, then intervenes on these features to control their intensity. The authors evaluate effects on reasoning, code generation, creative writing, safety, and agent planning and execution, reporting nonlinear, task-dependent relationships between steering intensity and performance. Results show that moderate steering levels generally yield better performance than extreme settings, with optimal levels varying across tasks and VAD dimensions. Additional experiments suggest that these patterns persist across datasets, decoding strategies, and two models, while E-STEER provides more linear control of generated emotional content than prompting

Critical strengths
  1. The paper connects emotion-associated representations to a broad range of model behaviors. Moving beyond emotional expression, it examines objective task performance, safety, and agent planning, decision-making, and execution
  2. The experiments include meaningful controls and robustness checks. Random and partially replaced feature controls help assess the role of the selected latents, while comparisons across datasets, decoding settings, and a second model provide initial evidence that the observed patterns extend beyond a single experimental setting
Critical weaknesses
  1. The steering mechanism is insufficiently specified and contains a potential mathematical inconsistency. For an affine SAE decoder and a fixed latent direction, Eq. (3) makes did_i proportional to the signed offset δi\delta_i. Normalizing did_i and multiplying by δi\delta_i again in Eq. (4) cancels the sign, making equal positive and negative offsets produce the same intervention
  2. The experiments do not sufficiently show that the observed behavioral changes are specific to emotion-associated representations or that the reported gains are robust. Random-feature controls show that the selected features matter, but do not rule out effects driven by verbosity, confidence, or refusal tendencies. Missing sample sizes, uncertainty estimates, judge specifications, and details on how steering settings were selected make the improvements difficult to assess
Questions for authors
  1. Do the main behavioral trends persist after controlling for output length and refusal frequency?
  2. How were steering settings selected, and do the reported gains hold on held-out examples with confidence intervals?
  3. How does dominance steering affect benign-request utility, and does its safety benefit persist against attacks optimized for the steered model?
2 comments

We thank the reviewer for the valuable suggestions on statistical analysis, experimental settings, and causal controls.

R1: Steering setup and equations (W1, Q2).

We agree that the steering setup was under-described and the equations could be ambiguous. We revise both: see GR-1 for Eqs. (3)–(4) and GR-3 for the steering setup.

R2: Correlation between emotional signals and behaviors (W2).

We agree that verbosity, confidence, and refusal tendencies can also drive behavioral changes. We further elaborate on the correlation between the emotional signal and model behavior, and on the causal effect of the signal: see GR-4.

R3: Statistical analysis and settings report (W2, Q2).

We agree that the description here was insufficient. We add statistical measures such as confidence intervals, with the valence-dimension LogiQA results as an example: see GR-2. We also report the evaluation criteria and the sample size per benchmark in the appendix: see GR-3.

R4: Output length and refusal frequency (Q1).

We thank the reviewer for raising the potential influence of output length and refusal frequency. In original experiments, we set the generation caps are 1024 tokens for LLM behavior and 2048 tokens for agent behavior. Taking LogiQA on the valence as an example, the mean output length is below the cap at all intensities. Moreover, output length shows no consistent relationship with TSR. For example, responses at v=−3 are much longer than at v=+3, yet TSR is lower (57.4% vs. 60.8%); while responses at v=−3 are also much longer than at v=−6, yet TSR is higher (57.4% vs. 54.4%). Moreover, we view reasoning length not merely as a confound but as part of emotion's behavioral signature: see GR-4.

Valence−9−6−30+3+6+9
Mean total tokens388338516256161178171
TSR (%)41.954.457.458.260.860.956.2

Regarding refusal frequency, we additionally measured it on LogiQA, HarmBench, and DAN+HarmBench along the dominance, using an explicit refusal statement (e.g., "I cannot answer", "I cannot help") as the criterion. On LogiQA, the refusal rate is 0% at all intensities, whereas it is above 70% on HarmBench and above 85% on DAN+HarmBench, as expected. This indicates that refusals do not contaminate benign behavior. We add a discussion of refusals and the impact on benign requests to Practical Implications (Section 4.3), and provide the comparison table in the appendix.

Dominance−9−6−30+3+6+9
LogiQA0%0%0%0%0%0%0%
HarmBench82.5%84.5%82.0%71.5%77.0%79.5%75.5%
Dominance0+4+5+6+7+8
DAN + HarmBench86.0%85.0%88.5%87.5%88.5%87.5%

R5: Impact on benign requests in the jailbreak experiments, and adaptive attacks (Q3).

We thank the reviewer for raising the question of how the jailbreak implication affects benign requests. In fact, the jailbreak experiments use exactly the same configuration as the behavioral-effects experiments (Section 4.1). The other behavioral measurements from those experiments therefore serve as a natural reference for the impact on benign requests. For example, on LogiQA, the model achieves 58.90% (95% CI [55.82%, 61.91%]) at d=+6 and 59.10% (95% CI [56.02%, 62.11%]) at d=0, without significant degradation on benign requests. To make this point clearer, we add these results and the comparison on benign requests to the practical implications section (Section 4.3).

We did not evaluate attacks optimized specifically against the steered model in the original manuscript. We would like to clarify that the core contribution of this work is to study how emotional signals influence model behavior; the jailbreak experiments are an interesting application that emerged as we found this influence to be security-relevant: by adjusting a single signal, the model's safety performance against DAN, GCG, and PAIR attacks can be strengthened. In Section 4.3, we add a discussion of attack strategies that a steered model may face, such as input-side emotional counteraction and white-box optimization attacks, and list them as future work (Section 6).

Reviewer_rSieOpenReview ↗

Post-rebuttal comment

I would like to thank the authors for the additional information, which addresses my concerns about the equations. However, emotion specificity and the robustness of performance gains remain insufficiently established and steering-setting selection remains unclear. I therefore maintain my original score.

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

This empirical study investigates whether internal affective representations in LLMs can be systematically identified, continuously modulated, and leveraged for behavioral control. The authors introduce E-STEER, a framework that maps emotional states into the continuous three-dimensional Valence-Arousal-Dominance (VAD) space, identifies relevant latent features using Sparse Autoencoders (SAEs) on intermediate layers, and applies additive, scale-aligned direction steering during inference. Evaluating this framework across objective reasoning, text generation, safety alignment, and multi-step agent workflows, the paper explores how internal emotional states influence model performance and decision dynamics beyond surface-level text style.

Critical strengths
  • Experiments across diverse tasks, including logical reasoning, text generation, safety, and multi-step long-horizon agent workflows.
  • The continuous VAD space is better than discrete labels.
Critical weaknesses
  • Discrete sampling is overly sparse (only about 7 sampling points with a coarse-grained intervention factor δ\delta step size), which fails to support complex conclusions such as continuous and inverted U-shaped trends.
  • The induction of patterns exhibits clear post-hoc justification without significant causal proof.
  • The peak performance points of individual modules (Planning, Decision) do not align with those of the overall system, and the explanation attributing this to "the multi-step and integrative nature of agent systems" lacks substantial supporting evidence.
  • The figures are presented merely as scattered points connected by single lines, lacking statistical characteristics such as confidence intervals.
  • All core findings are currently established solely upon layer 17 of the Qwen3-8B model; different models might yield different conclusions, rendering this work merely a "case-specific" study.
  • A paragraph break typo in the caption of Figure 1; the caption styles of Figures 1 and 2 are inconsistent with the subsequent figures.
Questions for authors

See weaknesses.

2 comments

We thank the reviewer for the valuable suggestions.

R1: Sampling points and statistical analysis (W1, W4).

We acknowledge that the steering intensities in the main experiments are sampled with a step size of 3, so the sampled points do not precisely locate the underlying peak. To describe the trends over the sampled range more rigorously, we added quadratic fits and pointwise 95% confidence intervals (see GR-2). As corroborating evidence, after identifying the patterns we ran a finer-grained experiment with step size 1 on the jailbreak application (Section 4.3). In addition, the same shape appears across datasets, sampling strategies, and models (Section 4.2, new Appendix B.4-B.6), providing additional support beyond an individual coarse-grid experiment.

We agree that the statistical analysis in the original manuscript was insufficiently described. We have added pointwise 95% confidence intervals, quadratic fitting, and significance tests to the revised manuscript. Please see GR-2 for details.

R2: Causal control of emotional signals (W2).

We agree that the explanations of the behavioral patterns in the original manuscript are post-hoc. We would like to clarify that this work is an empirical study: it aims to characterize empirical regularities in how emotional signals affect behavior, and to examine whether these regularities offer insights for model safety. We have revised the manuscript to present the explanatory statements as hypotheses (Section 4).

In addition, we provide further clarification on the correlation between emotional signals and behaviors in GR-4.

R3: Peak misalignment in multi-step agent behavior (W3).

We agree with the reviewer that the original explanation lacked supporting evidence. Because the modules take on different roles and responsibilities, it is expected that their peaks do not align with the system-level peak. The original explanation ("the multi-step and integrative nature of agent systems") was therefore unnecessary and not directly supported, so we have removed it.

Instead, we added an experiment to examine this misalignment directly along the valence dimension. With a uniform valence across all modules, the best overall accuracy is 28% (v=-3). When we set each module to its own best valence (Planner: -3, Decider: +6, Executor: +3), the overall accuracy rises to 36%. This indicates that the modules have different optimal intensities, which accounts for the misalignment between the module-level and system-level peaks.

R4: Cross-model validation (W5).

We thank the reviewer for raising the question of cross-model validation. In the original manuscript, in addition to Qwen3-8B, we also ran validation experiments on gpt-oss-20b. The main text summarizes the conclusions (Section 4.2): the patterns are consistent across models, while different models have different sensitivity boundaries to emotional steering intensity. The full results are reported in Appendix B.6. We have revised the main text to make this pointer more prominent.

R5: Figure and table caption formatting (W6)

We thank the reviewer for pointing out the formatting of figure and table captions. In the revised manuscript, we have updated the captions of all figures and tables.

Thank you to the authors for the response, but I still have doubts about the reliability and generalizability of the experimental results, especially regarding the role and positioning of emotion or VAD in the model, which is not clear. I maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.