acceptodds
Under review as a conference paper at ICLR 2027

Valence Asymmetries In Language Models: Representation, Behaviour, And Self-Report

Abstract

Emotion-related representations in large language models can influence be- haviour, but whether their internal geometry, behavioural effects, and self-report reflect the same information remains unclear. We investigate these relationships across 171 emotions and 15 models, finding a near-planar valence–arousal ge- ometry that emerges with depth and persists through post-training, with valence consistently more decodable than arousal. Valence also shapes safety-related be- haviour: it couples strongly with refusal (ρ = −0.82), while 83% of multi-turn attack trajectories exhibit negative-valence drift. However, internal representation does not guarantee faithful self-report. Despite near-perfect decoding of induced emotions, models exhibit pronounced valence-dependent reporting asymmetries, which become more apparent under competing emotional inductions, where inter- nal representations and reported states frequently diverge. Together, our findings reveal valence as a recurring organising dimension of affective representations while highlighting fundamental distinctions between what models represent, how those representations influence behaviour, and what models report about their in- ternal states.

Primary area:
interpretability and explainable AI
Keywords:
valence–arousal circumplex, emotion representations, mechanistic interpretability, representation engineering, activation steering, refusal direction, AI safety

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_F4RsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper extracts 171 emotion vectors from synthetic stories, analyzes their structure across 15 models or training checkpoints, and investigates the relationships between emotion representations and refusal behaviour, multi-turn attacks, self-report, and evaluation awareness. An important finding is that, although external probes can identify the injected emotions with near-perfect accuracy, models do not always select the corresponding emotion labels in self-report. In the competing emotional induction experiments, activation steering has a greater influence on the measured internal representations, while model self-reports tend to follow textual instructions. In addition, the authors find a strong relationship between valence and refusal behaviour. However, in the evaluation-awareness task, strong representational alignment does not necessarily imply a corresponding causal dependence.

Critical strengths
  1. The paper is well motivated. Studying these questions within a unified framework of emotion representations may help us better understand the reliability of model self-report. The experiments provide good support for the conclusion that information recovered from internal representations by external probes is not necessarily reflected in the model's verbal self-report.
  2. The authors investigate emotion representations across multiple model families and analyze their relationships with different downstream behaviours. The paper combines several methods, including linear probing, activation steering, cosine-matched controls, behavioral evaluation, and representational ablation.
Critical weaknesses
  1. One of the paper's main conclusions is that models can accurately represent injected emotion information internally but tend to select positive emotion labels in self-report. However, the results show that models already exhibit a strong preference for positive labels even without emotion injection.For example, in Qwen3.5-9B, the positive- and negative-valence target-choice rates are 93.1% and 1.9%, respectively, without steering. After steering, these rates become 91.9% and 5.9%. Although the authors have discussed this issue, it remains unclear how much the pre-existing label preference and the changes induced by emotion steering each contribute to the observed valence asymmetry. Therefore, I think the authors should further clarify to what extent the baseline preference affects the interpretation of the self-report asymmetry.
  2. The paper finds a significant relationship between valence and refusal behaviour, and the authors provide additional evidence through cosine-matched controls. However, Table 1 in the appendix shows that, after accounting for the geometric relationship between emotion vectors and the refusal direction, adding valence and arousal provides statistically significant additional explanatory power in only three of the seven models. This suggests that the relationship between valence and refusal behaviour is relatively consistent, but the evidence for how much additional explanatory power valence and arousal provide after controlling for simple geometric factors is less consistent across models. Furthermore, cosine-matched controls cannot fully rule out the influence of other features carried by the emotion vectors.
Questions for authors
  1. Could you explain the positive- and negative-valence reporting differences across models after accounting for pre-existing label preferences, and explain how these changes affect the interpretation of Valence Readout Asymmetry?
  2. Could you further clarify the independent role of valence in refusal behaviour, for example, removing the valence component while keeping other conditions unchanged, and comparing the behavioural effects before and after removal?
2 comments
AuthorsOpenReview ↗

Response to Reviewer F4Rs

  1. We thank Reviewer F4Rs for highlighting the need to clarify how pre-existing label preferences and steering-induced changes contribute to reporting asymmetry across models and affect the interpretation of VRA. We added a post hoc analysis in Appendix K, after “Interpreting the readout gap.” With positive/negative target-choice rates R+,R- and reporting gap A=R+-R-, we decompose A1=A0+(R+-R-); 0/1 denote unsteered/steered conditions and R denotes the change from baseline.

All seven models favour positive labels without steering. Additive steering changes choice rates as follows (percentage points):

ModelPositive ΔR₊Negative ΔR₋
Qwen 2B+4.38−2.50
Qwen 4B−4.06+2.81
Qwen 9B−1.25+4.06
Qwen 27B−15.00+17.50
Llama 8B−22.19+13.44
Gemma E4B+7.19+24.69
Gemma 31B−23.44+13.44

Five models increase negative and decrease positive endorsement; Gemma E4B increases both, especially negative endorsement. Qwen2B moves oppositely, but its gap-change interval includes zero. Addition narrows the gap in six models, also on the original 18 VRA-matched pairs. For Qwen9B, the gap falls from 91.25 to 85.94 points (−5.31). Bootstrap intervals are in Appendix K.

Thus, the large Qwen9B asymmetry already exists before injection, which partially reduces it. Across models, baseline preference and intervention response make distinct contributions. VRA describes the remaining reporting disparity alongside recoverable injected identity; it is not a baseline-adjusted causal effect of injection.

  1. We thank the reviewer for distinguishing a consistent valence–refusal association from a valence-specific causal effect. Our claim is the former: valence orders the effect of emotion steering on refusal. We have revised the few sentences that implied the latter, such as "gates" and "acts through". The evidence for the ordering claim is as follows.
  1. Ordering. Valence predicts the steering effect in 7/7 models (r = 0.36–0.73, p < 10⁻⁵). It replicates in Qwen3.5-2B, -4B and -9B on 24 emotions held out from axis fitting (r = 0.43, 0.72, 0.56; the bootstrap CIs exclude zero for 4B and 9B).
  2. Beyond geometric overlap. Emotion directions exceed cosine-matched random directions in 6/7 models (d = 0.79–1.86). The reviewer notes that the joint valence–arousal test does not isolate valence, so we added a valence-only test given geometry and arousal; it is significant in 4/7 models. While doing so, we found that the Table 1 F-tests used n = 48 degrees of freedom for regressions fitted on all 171 emotions. With the correct degrees of freedom, the joint test is significant in 6/7 models rather than 3/7. Refits on the 48-emotion subset (4/7 and 2/7) are reported as a sensitivity analysis.
  3. Component removal, as requested. We removed the fitted valence axis from each held-out emotion direction in three Qwen3.5 sizes (20,844 graded generations). The controls are norm-matched, random, arousal-removal and refusal-projection-preserving edits.
    • Removal does not detectably weaken the valence ordering: all multiplicity-adjusted CIs span zero.
    • However, valence remains decodable from the edited directions (held-out R² = 0.745–0.789, against 0.807–0.839 before removal), and the ordering persists.
    • Valence is therefore encoded redundantly. Single-axis removal tests one readout direction rather than valence itself, and we do not read it as evidence for or against a valence-specific causal role.

We added a valence-component-removal experiment (§4.2, Appendix “Valence-component removal in refusal steering”): on 24 held-out emotions in Qwen3.5-2B/4B/9B, the valence ordering of refusal replicates, and removing the fitted valence axis does not detectably weaken it while valence remains decodable (held-out R² = 0.745–0.789), indicating that valence is encoded redundantly rather than in a single removable direction.

AuthorsOpenReview ↗

Gentle reminder before the discussion closes

Thank you again for your constructive feedback. With less than one day remaining in the discussion period, we would appreciate any further thoughts on our response. To briefly summarize the key revisions and additional experiments:

  • Clarification of VRA: We separated pre-existing label preferences from steering-induced changes, clarifying that VRA measures the remaining reporting disparity rather than a baseline-adjusted causal effect.
  • Statistical analysis: We added valence-only statistical tests and corrected the regression degrees of freedom.
  • Valence-component removal experiment: We conducted the requested experiment with matched controls. Removing the fitted valence axis did not detectably weaken the behavioural ordering, while valence remained decodable.

Accordingly, we revised our conclusions to distinguish valence's predictive relationship with refusal from evidence of causal necessity. We would be happy to clarify any remaining concerns and appreciate your consideration of these revisions.

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

This paper investigates the relationship between emotion representations, model behavior, and self-report in large language models. The authors extract representations of 171 emotion concepts across 15 models and examine their organization along valence and arousal dimensions. Through activation steering, refusal experiments, multi-turn adversarial interactions, and introspective evaluations, they investigate whether these representations influence model behavior and whether internal states correspond to reported emotions. The results suggest that valence is a prominent organizing dimension, influences refusal-related behavior, and exhibits systematic discrepancies between internal representations and conditional self-report.

Critical strengths

1. Important research question and broad evaluation: The paper addresses an important question about whether emotional representations in LLMs correspond to their behavior and self-report. Evaluating affective geometry across 15 models from different families strengthens the analysis beyond a single-model study. The finding that valence and arousal representations become more clearly organized through early-to-middle network layers is particularly interesting and consistent with the broader understanding that Transformer layers progressively develop more contextualized representations. The analysis of these representations across model depth and post-training is informative.

Critical weaknesses

1. Insufficient robustness validation of emotion-vector construction: The paper adopts emotion-vector extraction from Emotion Concepts and their Function in a Large Language Model (Sofroniew et al., 2026), whose authors acknowledge potential topic-specific confounding in synthetic-story representations. Although the original work provides external-context validation, the present study does not establish whether its cross-model geometry and downstream steering results remain stable when vectors are independently constructed from topic-disjoint stories. Since these vectors underpin the main experiments, their robustness across narrative contexts and extraction settings is important for interpreting the findings.

Questions for authors

Q1. Emotion corpus construction and provenance: Could the authors please clarify whether they reuse the complete synthetic-story corpus from Emotion Concepts and their Function in a Large Language Model (Sofroniew et al., 2026) or generate a new corpus? Please specify the exact number of stories per emotion and whether the original story-generation and validation procedures were retained.

Q2. Emotion-vector robustness and topic confounding: Sofroniew et al. (2026) acknowledge that emotion vectors derived from synthetic stories may contain topic-specific confounds. Have the authors examined whether independently constructed vectors from topic-disjoint story subsets produce similar directions, valence-arousal geometry, and steering effects? For example, would sadness vectors extracted from stories about a child breaking a favorite toy versus an adult grieving the loss of a close friend produce comparable results? Additionally, how sensitive are these findings to the fixed token-50 pooling strategy, given that emotional content may emerge at different positions within a story?

2 comments
AuthorsOpenReview ↗

Response to Reviewer MJza

We thank the reviewer for these questions. In the updated submission, We have added construction and validation details to Sections 3.1–3.2 and Appendix N. Sections 4.1 and 4.4 now briefly report pooling robustness, with the full comparison in Appendix O.

Q1: Emotion corpus construction and provenance

Construction and extraction (Sections 3.1–3.2; Appendix N.1–N.3). We generated 6,510 new stories across 171 emotions with Qwen3.5-4B. We followed Sofroniew et al.'s prompt, topic list, and extraction protocol. For each emotion, we sampled ten of their 100 topics and requested four stories per topic. The prompt conveys emotion indirectly and forbids the target word and direct synonyms. We screened short segments and obvious target-word/stem matches, and manually inspected stories for 30 common concepts. Appendix N.2 gives three examples.

We pool activations from token 50 onward, centre each emotion mean against the mean across emotions, and remove the leading components explaining 50% of neutral activation variance. The neutral corpus contains 200 factual passages.

Validation (Section 3.2; Appendix N.4–N.5). We examine 20 Qwen emotion directions on 6,941 human-written documents from GoEmotions (Demszky et al., 2020) and EmpatheticDialogues (Rashkin et al., 2019). Figures 22–23 show emotion-relevant, high-activation texts outside the construction corpus. A complementary logit-lens check finds that the excited and relieved directions upweight their corresponding labels by 0.058 and 0.078 logit units; the sad direction upweights mourn by 0.061 (Figure 24). Together, these checks support semantic alignment beyond the synthetic stories. The examples are selected activation maxima, so they do not estimate classification accuracy.

Q2: Emotion-vector robustness and topic confounding

Construction controls. Since we are using multiple (10) topics per emotion and neutral-subspace removal, checking if a topic-specific confound exists makes less sense as our method reduces reliance on a single narrative setting or shared non-emotional features thus reducing any topic-specific confounding. We do agree that some dataset confounds could remain due to the particular dataset we constructed as ackowledged by Sofroniew et al. also in their paper.

Pooling sensitivity (Sections 4.1 and 4.4; Appendix O). We compare all tokens, token 25 onward, token 50 onward, and the final half. All four conditions use the same 40-emotion subset from Section 3.5, with neutral components and evaluation procedures fixed. Reconstructing probes on this subset can change exact values relative to the main study's 171-emotion construction. The qualitative conclusions remain the same.

The table below reproduces Appendix O's comparison (Table 10):

ModelAll tokens25+50+Final half
Valence/arousal held-out R², layer 16
Qwen3.5-4B0.881/0.5380.876/0.5590.873/0.5820.858/0.579
Llama-3.1-8B-Instruct0.849/0.5660.833/0.5600.842/0.5780.816/0.565
Valence readout asymmetry (VRA)
Qwen3.5-4B1.5431.5721.5431.579
Llama-3.1-8B-Instruct0.4510.3680.4860.486

Valence remains more decodable than arousal in every condition. Direction agreement is also high. At layer 20, alternative-window directions have mean cosines of 0.972–0.986 for Qwen and 0.963–0.984 for Llama with the 50+ directions (Figure 25).

The reporting asymmetry persists across windows. VRA stays positive, while identity decoding remains near ceiling (99.9–100% mean chance-corrected recall). Thus, changing the pooling window preserves both the affective ordering and the larger reporting shortfall for negative labels.

References

  • Demszky, D., et al. (2020). GoEmotions: A Dataset of Fine-Grained Emotions. arXiv:2005.00547.
  • Rashkin, H., Smith, E. M., Li, M., & Boureau, Y.-L. (2019). Towards Empathetic Open-domain Conversation Models: A New Benchmark and Dataset. ACL, 5370–5381.

I thank the authors for thoroughly answering my questions. The new details mostly confirm the robustness of their findings, and thus I recommend they be explicitly discussed in the final paper if accepted. I retain my original score.

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

The paper presents an empirical study of emotion-related representations (now often called emotion vectors). In a probing study of multiple models across a previously established corpus of 171 emotions, they identify a planar „affect geometry” of the valence-arousal emotion subspace in representations. They show how these concepts are related to behavior, such as refusal, and how self-reporting at the output and probing of internal representations diverge.

I think this manuscript has the ingredients of a nice paper. There are a series of interesting results, and the experiments performed are extensive. I would encourage the authors to revise their draft and streamline the presentation/main storyline. I would be ready to revise my score in this case.

Critical strengths
  • The paper is well-written, clear, and easy to follow
  • The contribution is an extensive empirical study across many conditions, probes, and research questions. The paper provides a deep insight into the geometry of valence and arousal directions within the activation space.
  • The experiment sections contain multiple findings, backed up by sufficient statistical power (i.e., by being identified across multiple models)
  • The paper includes some well-crafted figures that provide strong visual intuition for the described concepts.
  • The appendix is extensive, containing a lot of details and additional results.
  • The chosen experiments, metrics, and methods are thoughtful and seem appropriate for the research questions.
Critical weaknesses
  • The paper reads a bit unfocused, containing a large set of experimental results and probing methods, but leaves it up to the reader to come up with a full storyline. This makes it a bit more difficult to come away with a clear take-home message. I think some curation of the performed experiments would be helpful
  • In general, the paper hides relatively modest novelty (compared to existing work like (Sofroniew et al.( and (Sun et al.)) behind a lot of content
  • I think providing source code would really strengthen the contribution, as there are a series of implemented probes & experiments.
Questions for authors
2 comments
AuthorsOpenReview ↗

Response to Reviewer erwF

We thank the reviewer for the feedback and agree with the fact that the paper needs some revision to streamline the main story. We have tried to address these concerns to our best and welcome any new feedback. A new version of the manuscript with the changes highlighted has been uploaded. The summary of the changes are follows:

1. Focus, curation, and storyline

Thank you for identifying the need for a clearer storyline. We agree that the original presentation gave too many experiments similar prominence and did not sufficiently explain how one motivated the next. We have revised the abstract, introduction, methodology, results, and discussion around one question: How do emotion-related representations, their behavioural influence, and self-report correspond, and where do they diverge?

The main text now follows three connected steps:

  1. Establish a common coordinate. The geometry analysis identifies valence as a shared coordinate for comparing emotion representations across models. We present this as the foundation for the subsequent analyses.
  2. Test behavioural influence and its limits. The refusal experiments ask whether this coordinate predicts intervention effects beyond initial geometric overlap. We distinguish refusal suppression from verified harmful completion. The multi-turn experiments extend this analysis to accumulating adversarial context, while making clear that affective drift does not directly rank attack success.
  3. Compare recoverable information with reporting. The introspection experiments ask whether injected emotion identity that is recoverable from activations is also reflected in conditional self-report. Competing emotional inductions then test whether the activation readout and report follow the same source when textual and activation-level cues disagree.

We have added explicit motivations and transitions between these steps. We have also curated the main text: the detailed multi-turn summary and secondary introspection-control overview now appear in their respective appendices. The evaluation-awareness protocol and results are collected in the appendix as a supporting comparison showing that strong valence alignment need not imply dependence on the valence component. The main text retains the unsteered reporting baseline, external-recognition comparison, and relevant interpretive limits.

2. Source Code for Reproducibility

We apologize that source code was not provided in the submission. Considering the fact that it is necessary for reproducing our experiments and important for other researchers to build upon, we have uploaded the source code for all our experiments in the supplementary material.

I thank the authors for their comprehensive revision. I think the streamlined presentation helps the storyline, and i appreciate the additional results. I have updated my score accordingly.

As the core novelty remains somewhat limited, i am not ready to further raise my score, but the paper is a well-executed investigation of the emotion representations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.