acceptodds
Under review as a conference paper at ICLR 2027

Characterizing the Geometric and Temporal Organization of Emotion in Neural Representation Spaces

Abstract

Emotion is encoded in the internal representations of language models, yet the principles governing its organization remain poorly understood. This raises a fundamental question: do emotional concepts form a structured geometry in representation space, and how does that structure evolve as emotional states unfold through conversation? We address these questions by modeling categorical, continuous, and temporal aspects of emotion through three representation forms, State, Continuous, and Dynamic, using BERT, DistilBERT, MiniLM, and RoBERTa. This choice of encoder-only Transformers provides contextualized representations while enabling comparison of emotional organization across architectures. Applying these representations to emotion classification and conversational datasets, we find that emotional concepts are broadly accessible through linear decoding. Yet this accessibility does not imply an unstructured representation. Emotional concepts concentrate in compact, structured subspaces with lower intrinsic dimensionality than the ambient representation space. The resulting organization varies with the representation objective, exposing different computational properties while preserving core geometric patterns across architectures. This organization also extends beyond individual emotional states into their evolution through dialogue. Across conversational trajectories, emotional representations exhibit reproducible temporal organization, with neighboring states showing predictive persistence and systematic changes in representation dynamics around emotional transitions. Together, these results establish a unified framework for characterizing emotion, linking the geometry of emotional states to representation learning and their evolution over time within conversational context.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Affective Computing, Emotion Representation, Latent Space, Neural Representation Spaces, Temporal Dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_oYxYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

It presents an empirical study of how Transformer (language) models represent emotional information, examining its accessibility, latent geometry, and evolution across dialogue turns. It compares three representations: State, Continuous, and Dynamic within the framework across four pretrained Transformer encoders and five affective datasets. The analysis combines emotion classification, regression, forecasting, and controllable editing with linear probing, intrinsic dimensionality estimation, manifold analysis, and temporal trajectory measurements. The results suggest that emotional information is somehow linearly accessible, occupies compact and locally structured latent spaces, and exhibits representation-dependent strengths. It also reports complementary information across the three representations and some evidence of temporal organization in dialogue. This study mainly focused on the systematic characterization of these relationships instead of a new emotion-recognition architecture.

Critical strengths
  • It brings together emotional accessibility, latent geometry, and dialogue dynamics within the same framework, comparing three affective representations across four Transformer models and five datasets. This provides a broader understanding of emotion representations beyond conventional task-specific performance comparisons.

  • It examines intrinsic dimensionality, manifold geometry, temporal organization, and complementary information across representations. The space dimensionality checks, capacity-controlled complementarity analysis, and temporal shuffling controls provide useful insights into how emotional information is organized and retained in language-model representations.

Critical weaknesses
  • Several key findings, such as linear decodability, low-dimensional emotional structure, and latent emotion dynamics, have already been investigated in recently published work, particularly the included references Reichman et al. (ICLR 2026) and Saini et al. (EMNLP 2026). Although the unified analysis is valuable, it does not establish what fundamentally new scientific understanding emerges from combining these perspectives. This limits my assessment of its originality and significance.

  • Temporal structure varies considerably across datasets, and the strongest chronological patterns may reflect dialogue conventions or semantic continuity instead of emotion-specific dynamics. Likewise, the task-dependent advantages of State, Continuous, and Dynamic representations may partly reflect their different learning objectives. Such limitations make some broader conclusions less convincing despite the extensive empirical analysis.

Questions for authors
  • Please clarify what genuinely new scientific insights emerge beyond Reichman et al. (ICLR 2026) and Saini et al. (EMNLP 2026)? I am particularly interested in what the joint analysis of State, Continuous, and Dynamic representations establishes that earlier studies did not.

  • The three given representations are learned using different objectives; how do we distinguish their inherent geometric properties from differences induced by task-specific supervision? Does the existing native-space analysis support the same conclusions?

  • Given that temporal patterns vary across datasets, what evidence supports interpreting them as emotion-specific dynamics rather than topic continuity, speaker effects, or general dialogue structure? How far do the existing controls address these alternatives?

  • It would be nice to know whether the capacity-corrected complementarity analysis genuinely establishes distinct emotional information across the three representations, rather than improvements from combining differently trained representations. Could you clarify what existing evidence supports this interpretation?

  • Could you clarify how the proposed geometric findings differ from Yamauchi and Aizawa (ACL 2026) and Choi and Weber (arXiv 2026), particularly regarding low-dimensional affective structure, linear accessibility, and the relationship between geometry and downstream performance?

  • Similarly, Maheswaran and Desarkar (EACL 2026) also investigate common emotional representations across models and datasets. What new scientific insights arise specifically from jointly comparing State, Continuous, and Dynamic representations beyond the findings of this and other cited studies?

  • By considering Hewitt and Liang (EMNLP 2019), could you clarify whether existing probing controls sufficiently distinguish emotional information encoded in the pretrained representations from information learned through supervised probes?

7 comments
AuthorsOpenReview ↗

Response to Weakness(W) 1 & 2 and Question(Q) 1

We thank the reviewer for the thoughtful and detailed assessment. The questions regarding emotion-specific dynamics, representational complementarity, and the distinction from prior work prompted several additional experiments and a reassessment of our original interpretations. We introduced a three-seed ensemble control, additional tests of dialogue-position effects, and Hewitt and Liang (2019) probing controls. These analyses have helped us distinguish inherited representational properties from those induced by supervision, while narrowing claims that were not sufficiently supported. All corresponding changes have been incorporated into the revised manuscript, with a highlighted version provided in the supplementary material.

W1 / Q1. What is new beyond Reichman et al. (ICLR 2026) and Saini et al. (EMNLP 2026)?

We agree that linear decodability and low-dimensional emotional structure have been established in prior work and no longer present them as novel findings. Reichman et al. characterize a low-dimensional, steerable emotional manifold in LLM hidden states, while Saini et al. study emotional trajectories and controlled transitions. Our contribution is to compare objectives grounded in distinct theories of emotion using the same encoders and shared baselines. This comparison reveals four findings:

  1. Geometry is induced rather than inherited. The pretrained encoder is identical across objectives, with mean ID 18.6 across four encoders and five datasets. Training reduces ID by 40–66%, depending on the objective (State: 6.4; Dynamic: 8.1; Continuous: 11.2), whereas a randomly initialized projector retains an ID of 17.4. Continuous has higher ID than State in all 20 encoder–dataset settings, with non-overlapping confidence intervals.

  2. Compactness is not emotion-specific, but emotional content is. A matched model trained on non-affective dialogue acts achieves comparable compactness (ID 6.6–9.9) to the emotion classifiers (5.2–8.6), whereas Continuous representations remain higher-dimensional (8.1–15.8). However, at comparable ID, emotion-trained representations decode emotion more accurately (macro-F1 0.345 vs. 0.220). Thus, supervision determines much of the geometry, while the prediction target influences the information accessible within it.

  3. Compression does not necessarily imply greater utility. Continuous, the least compressed objective, transfers best to held-out IEMOCAP. Moreover, pretrained BERT [CLS] representations achieve higher IEMOCAP emotion F1 (0.45) than any fine-tuned BERT latent (0.32–0.41).

  4. Dialogue regularity reflects position rather than affect. The observed drift is shared across objectives, persists when considering neutral utterances alone and after controlling for emotion and dialogue act, and shows little alignment with valence. This provides a direct test of whether the observed temporal structure is genuinely affective, which a single-objective analysis cannot establish.

We have revised the Introduction to state these contributions explicitly.

W2. Temporal structure varies; task advantages may reflect the different objectives.

We agree that the original temporal interpretation required further scrutiny. Our additional analyses confirm that the DailyDialog effect primarily reflects discourse position rather than emotional dynamics, and we have revised the corresponding claims accordingly. Regarding task performance, we view its dependence on the training objective as a central aspect of our investigation rather than a confound. The revised experiments directly quantify these differences against a shared pretrained baseline. The matched non-affective model further distinguishes the effects of supervision type from those of the emotional prediction target.

AuthorsOpenReview ↗

Response to Question(Q) 2 and 3

Q2. Inherent vs supervision-induced geometry; does the native-space analysis agree?

We thank the reviewer for highlighting the need to distinguish inherent encoder geometry from changes introduced by supervision. Our original native-space analysis used the fine-tuned encoder and therefore could not establish this distinction. We now include the pretrained encoder as the reference, evaluated on the same utterances using the same code.

The pretrained encoder has a mean ID of 18.6 across the 20 encoder–dataset settings, identical across objectives. The corresponding fine-tuned encoder IDs are 10.0 for State, 14.9 for Continuous, and 15.2 for Dynamic; the projected latent IDs are 6.4, 11.2, and 8.1, respectively. The pretrained baseline is consistent across alternative representations: mean pooling yields 19.0, PCA-128 yields 16.3, and the randomly initialized projector yields 17.4.

These results distinguish the inherited moderate dimensionality of pretrained representations from the objective-dependent compression induced by training. Continuous has higher ID than State in all 20 settings in both the fine-tuned encoder and projected latent spaces, with non-overlapping confidence intervals. Thus, the native-space analysis supports the same ordering, indicating that the projector further accentuates a difference already present in the fine-tuned encoder. IEMOCAP remains an exception: its pretrained representation is already low-dimensional (ID = 7.5), which we now explicitly identify as an inherited property.

Q3. Emotion-specific dynamics, or topic, speaker and dialogue structure? We thank the reviewer for prompting us to examine alternative interpretations of the observed temporal patterns. Our additional analyses indicate that the observed DailyDialog effect primarily reflects dialogue position rather than emotion-specific dynamics. We fitted the drift direction on training dialogues and evaluated it on held-out dialogues.

• Dialogue position: The drift direction predicts turn position on DailyDialog, with within-dialogue Spearman correlations of 0.36–0.47 across all 12 encoder–objective settings. Directions fitted on shuffled dialogues achieve only 0.01–0.08 (p = 0.005), the minimum attainable using 200 random directions).

• Emotion: The association persists when considering neutral utterances alone (rho = 0.34-0.44) and after controlling for emotion and dialogue act (rho = 0.23-0.37). Alignment with the valence axis is weak (|cos| <= 0.16).

• Topic and speaker: Since DailyDialog topic remains constant within each dialogue, the within-dialogue correlations control for topic. Speaker identity is also poorly decoded from the representations: on MELD, six-character classification achieves macro-F1 of 0.196, close to the 0.17 chance level. Moreover, alternating speakers in DailyDialog would produce an alternating rather than monotonic pattern.

• Objective and dataset: The effect is comparable across State, Continuous, and Dynamic (0.40, 0.44, and 0.40, respectively), indicating that it is not specific to the recurrent objective. The analysed Dynamic latents are per-utterance projections, with no recurrent processing at test time. The effect is also weaker on MELD (0.02–0.15) and IEMOCAP (0.07–0.21).

We have consequently revised the interpretation to describe this finding as positional drift and removed emotion-specific temporal claims from the abstract and conclusion.

AuthorsOpenReview ↗

Response to Question(Q) 4 to 6

Q4. Does RCS* show distinct information, or only gains from combining differently trained models?

We agree with the reviewer's distinction and thank them for identifying this limitation. RCS* controls probe width but does not independently control for the benefits of ensembling. To address this, we introduced a three-seed ensemble baseline: for each objective, we combined three independently trained BERT models using different seeds and evaluated them with the same RCS* procedure on identical test rows. Comparisons use paired, dialogue-clustered 95% confidence intervals.

The results show that combining State, Continuous, and Dynamic does not outperform the best same-objective three-seed ensemble on any dataset. The difference (three-objective ensemble minus best same-objective ensemble) is:

• GoEmotions: −0.060 [−0.092, −0.027]; the best same-objective ensemble is Continuous (+0.056 [0.030, 0.079]).

• MELD: −0.023 [−0.048, 0.005]; the best same-objective ensemble is State (+0.030 [0.007, 0.052]).

• DailyDialog: −0.021 [−0.064, 0.045]; the best same-objective ensemble is State (+0.056 [0.012, 0.094]).

• IEMOCAP: −0.059 [−0.117, −0.003]; the best same-objective ensemble is Dynamic (+0.126 [0.060, 0.197]).

The differences are significantly negative on GoEmotions and IEMOCAP, while the confidence intervals include zero on MELD and DailyDialog. Thus, the original complementarity result does not establish that combining objectives provides distinct emotional information; the observed gains are adequately explained by ensembling. We have corrected this interpretation in Section 4.6 and Appendices B.1 and F.5 and now report RCS* alongside the same-objective ensemble baseline and its recomputed values.

Q5. How do the findings differ from Yamauchi & Aizawa (ACL 2026) and Choi & Weber (arXiv 2026)?

We thank the reviewer for pointing us towards these relevant studies. Both are now discussed in the related work.

• Yamauchi and Aizawa impose circumplex geometry through hyperspherical contrastive learning and report a trade-off between psychologically grounded geometry and task performance. In contrast, we impose no geometric structure but examine what standard supervision objectives induce relative to pretrained representations. Our findings reveal a related trade-off: the most compressed representation is not necessarily the most transferable.

• Choi and Weber show that LLM emotion representations align with valence–arousal structure and are nonlinear but remain well approximated by linear probes. We similarly observe near-linear accessibility (mean tuned-probe gap of +0.006). However, our non-affective controls show comparable near-zero gaps for dialogue acts, speaker identity, and random control labels, suggesting that linear accessibility is not specific to emotion. Although a transferred valence axis is detectable in the full latent space (AUROC 0.71–0.97), its contribution to PC1–2 is no greater than that of a random direction. Thus, two-dimensional visualizations alone do not establish circumplex organization.

• Geometry and performance: Our additional transfer results also show that representational compression does not necessarily predict downstream utility, as discussed above.

Q6. What is new beyond Maheswaran & Desarkar (EACL 2026)?

We are grateful to the reviewer for drawing our attention to this relevant line of work. Maheswaran and Desarkar show that emotion reading directions in later layers of decoder LLMs are prompt dependent and interchangeable across datasets. Our work addresses a complementary question i.e. how different supervision objectives shape the representational spaces in which such directions are embedded. Across five datasets, the objective dependent ID ordering is consistent in all 20 encoder dataset settings. Furthermore, our non-affective controls distinguish compactness induced by categorical supervision from the emotional information accessible under emotion-targeted training. We have incorporated this distinction into the related-work discussion.

Reviewer_q1S5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper investigates the geometric and temporal organization of emotion in neural representation spaces. The authors construct three representation types—State, Continuous, and Dynamic—using four Transformer encoder backbones and evaluate them across multiple emotion datasets. Through probing, intrinsic dimensionality estimation, geometric analysis, and conversational trajectory analysis, the paper reports that emotional information is largely linearly accessible, occupies low-dimensional subspaces, and exhibits temporal regularities in conversational settings.

Critical strengths

(1) The paper addresses an interesting research question by examining the organization of emotional information beyond conventional downstream emotion recognition performance. Integrating information accessibility, latent geometry, and temporal analysis provides a broad perspective on affective representations.

(2) The experimental evaluation covers multiple datasets and encoder backbones, with complementary analyses including different probe complexities, native-space dimensionality measurements, and shuffled temporal controls.

Critical weaknesses

(1) Limited contribution and insufficient interpretation of the encoder and projection

The core experiments are limited to comparisons among the three representations proposed by the authors, without external baselines, and the reasons for the performance differences and their statistical significance are not sufficiently discussed. In particular, for the two-dimensional visualizations obtained through projection and PCA, both of which can lose information, the paper mostly describes the shapes of the plots without justifying how well they represent the original representations or what geometric meaning they have for downstream tasks. Moreover, the main text (Sec. 3) describes the encoder as frozen, whereas App. A states that it is fine-tuned after the first two epochs, and such inconsistencies reduce the reliability of the experimental protocol.

(2) Insufficient evidence for emotion-specific structure and temporal dynamics

The paper does not use non-affective controls to determine whether the observed low intrinsic dimensionality and geometric properties are specific to emotion or are general properties of language representation spaces (according to Sec. 4.2, intrinsic dimensionality is already low in the encoder space before projection). The temporal effects are also concentrated mainly on DailyDialog, and the authors themselves acknowledge that they may reflect discourse position or lexical cues (App. E.4), so the evidence is not sufficient to establish the temporal organization of emotion claimed in the title.

Questions for authors

(1) Emotion-specific geometry (Section 4.2, Tables 1 and 6). Given the low intrinsic dimensionality in the native encoder space and inconsistent geodesic advantages, what evidence distinguishes emotion-specific geometric structures from general properties of language representations?

(2) Encoder training protocol (Section 3, Figure 2, Appendix A). The main text describes frozen encoders, whereas Appendix A reports joint fine-tuning. Could the authors clarify the actual training protocol?

(3) Representation construction (Section 3.1, Figure 2, Appendix A). Could the authors clarify the architectures, parameter sharing, and training objectives of the three representation functions to ensure reproducibility?

(4) Statistical significance (Tables 1 and 2). Could the authors provide confidence intervals or significance tests to establish whether the reported nonlinearity gaps and representation–task differences are statistically meaningful?

(5) Temporal organization (Section 4.5, Appendix E.4). Given the DailyDialog-specific temporal effects and similar gains for State and Dynamic representations, how can these findings be distinguished from general discourse or lexical patterns?

5 comments
AuthorsOpenReview ↗

Response to Weakness(W) 1

We sincerely thank the reviewer for their insightful and thoughtful engagement with our work. Their observations have prompted us to strengthen the manuscript, sharpen our interpretations, and present our contributions with greater precision. The reviewer’s insightful observations on external baselines, non-affective controls, and consistency in the training protocol encouraged us to undertake additional experiments and revisit key methodological aspects, resulting in a more rigorous and comprehensive manuscript. All new analyses use the original splits, seeds, and code, which we verified by reproducing every published result exactly. All changes are incorporated into the revised manuscript, with a highlighted version provided in the supplementary material.

W1. No external baselines; no significance; 2-D projections; frozen vs fine-tuned.

We address the reviewer’s concerns through additional baseline comparisons, statistical testing, and methodological analyses: Baselines: We evaluated the same utterances, aligned by ID, in the pretrained encoder, PCA-128, and the randomly initialized 128-dimensional projector. Across four encoders and five datasets, mean ID was 18.6/19.0 for pretrained representations ([CLS]/mean pooling), 16.3/17.4 for PCA-128/random projections, and 6.4, 11.2, and 8.1 for projected State, Continuous, and Dynamic representations, respectively. The corresponding fine-tuned encoder IDs were 10.0, 14.9, and 15.2. MiniLM uses 384 dimensions; BERT, DistilBERT, and RoBERTa use 768. These results indicate that compression arises primarily through training, with its extent depending on the objective. IEMOCAP is an exception: it’s pretrained ID is already low (7.5) and changes relatively little after training. Significance: Appendix G.7 reports 95% bootstrap confidence intervals for all intrinsic dimensionality estimates and classification, regression, and forecasting scores (excluding editing). Resampling is performed at the dialogue level for MELD, DailyDialog, and IEMOCAP. Paired comparison procedures are detailed in our response to Q4.

2-D projections: We agree that the original discussion over-interpreted these visualizations. Restricting probes to PC1–2 reduces emotion F1 by 26–61%. An EmoBank-fitted valence axis has only 0.0–1.8% of its energy captured by PC1–2 (random direction: 1.6%) yet achieves AUROC values of 0.71–0.97 in the full latent space. We now label all 2-D visualizations as illustrative and base our conclusions on full-space analyses.

Frozen vs fine-tuned: We are grateful to the reviewer for highlighting this inconsistency, which prompted us to revisit and correct the relevant methodological description. The description in Section 3 did not accurately reflect the training protocol reported in Appendix A. In the actual implementation, the encoder is frozen for epochs 1–2 and fine-tuned during epochs 3–6. This fine-tuning has a substantial effect: BERT's encoder ID on GoEmotions decreases from 26.8 before training to 13.9 after State training. We have revised Section 3, Equation 1, the Figure 2 caption, and the Ethics statement to ensure that the training protocol is described consistently throughout the manuscript.

AuthorsOpenReview ↗

Response to Weakness(W) 2 and Question(Q) 1

W2. No non-affective controls; temporal effects concentrated in DailyDialog.

We are sincerely grateful to the reviewer for this insightful observation, which encouraged a deeper examination of whether the representational properties identified in our study are specific to emotional information or extend to non-affective tasks. We conducted three complementary control experiments to investigate this distinction: (1) non-affective probes on the same latent representations, covering dialogue acts (DailyDialog), speaker identity (MELD main characters and IEMOCAP gender), the Hewitt and Liang (2019) control task (random labels drawn from the emotion distribution for each unique utterance), and permuted labels; (2) pretrained encoder baselines; and (3) matched State models trained on DailyDialog dialogue acts and emotions, using the same 43,390 utterances and random seed.

The matched experiment shows that dialogue-act and emotion supervision produce comparable ID (7.53 vs. 8.09), while the Continuous objective produces higher-dimensional representations. The confidence intervals for these dimensionality comparisons do not overlap. Thus, compactness reflects categorical supervision rather than emotion alone.

However, the emotion-trained model achieves substantially higher emotion F1 than the dialogue-act-trained model on the same utterances (0.345 vs. 0.220). We quantify this distinction through selectivity, defined as the difference between a linear probe's macro-F1 on emotion and its macro-F1 on the control task, evaluated on the same test utterances. Across 4 encoders × 3 objectives × 4 datasets, using one probe per cell and seed 42, selectivity is positive in all 48 comparisons (mean = +0.207; lowest lower confidence bound = +0.037). Confidence intervals use paired bootstrap resampling (1,000 resamples), with dialogue-level resampling for conversational datasets.

Finally, the nonlinearity gap is close to zero across emotion (−0.022), dialogue acts (+0.001), speaker identity (+0.021), and the control task (+0.002). We therefore no longer interpret linear accessibility as emotion specific. Regarding temporal effects, we agree that their concentration in DailyDialog warrants further investigation. We address this through the additional analyses reported in our response to Q5.

Q1. What distinguishes emotion-specific geometry from general properties of language representations?

We sincerely thank the reviewer for raising this important distinction. The question prompted us to examine our findings more closely through additional controls and analyses, allowing us to distinguish the geometric effects of supervision from the emotion-specific information accessible within learned representations.

Our original description of the “native space” in Section 4.2 was insufficiently precise. It refers to the fine-tuned encoder rather than the pretrained encoder. The pretrained encoder has a mean ID of 18.6 and is identical across objectives, so it cannot account for differences between them. The ordering State < Continuous emerges only after training and holds across all 20 encoder–dataset settings, both in the fine-tuned encoder and projected latent space zz, where confidence intervals are non-overlapping in all 20 settings.

Our additional non-affective controls demonstrate that compactness is not unique to emotional representations but also arises from categorical supervision. A State model trained on non-affective dialogue acts has ID 6.6–9.9, comparable to emotion classifiers (5.2–8.6) and below Continuous (8.1–15.8). Similarly, linear accessibility is not exclusive to emotion: the nonlinearity gap is near zero for emotion, dialogue acts, speaker identity, and the control task (−0.022, +0.001, +0.021, and +0.002, respectively). We also revisited the original geodesic analysis and identified an implementation error that made the probe equivalent to Euclidean 1-NN. We have therefore withdrawn the local-structure advantage claim and included the corrected analysis in our response to Reviewer nxnq (W3).

To isolate the contribution of emotion-related supervision, we conducted a matched comparison using the same 43,390 DailyDialog utterances, architecture, and random seed. Despite comparable intrinsic dimensionality (ID = 8.09 vs. 7.53), the emotion-trained model achieves substantially higher emotion macro-F1 than the dialogue-act-trained model (0.345 vs. 0.220). This distinction is further supported across all 48 encoder–objective–dataset settings: emotion selectivity over the Hewitt and Liang control task is positive in every cell, with a mean difference of +0.207 and confidence intervals entirely above zero.

Section 4.2 has been revised to reflect this distinction of how representational geometry is shaped by the form of supervision, while the accessibility of emotional information is target specific. We are grateful to the reviewer for prompting this more nuanced examination of our findings.

AuthorsOpenReview ↗

Response to Question(Q) 2 to 5

Q2. What is the actual training protocol?

Each objective fine-tunes an independent copy of the pretrained encoder together with its corresponding prediction heads. The encoder is frozen for epochs 1–2 and jointly fine-tuned during epochs 3–6. Training uses AdamW (learning rate 2e-5, weight decay 0.01), a warm-up ratio of 0.1, gradient clipping of 1.0, and batch size 16 (two dialogues for Dynamic). We retain the checkpoint with the lowest validation loss, extract its representations, and train all probes on the resulting frozen latents. Section 3, Equation 1, the Figure 2 caption, and the Ethics statement have been revised to consistently describe this protocol.

Q3. Architectures, parameter sharing and objectives.

We summarize the training setup in Table 1 (Section 3.1), including the four encoder backbones, objective-specific supervision, and independent fine-tuning of each model without parameter sharing. All objectives use the same 128-dimensional projection architecture, while the prediction heads and loss functions differ according to the objective. For Dynamic, the GRU provides contextual supervision during training, whereas the analyzed latent remains the individual utterance projection. Full architectural details, training hyperparameters, and seed configurations are provided in Appendix A.

Q4. Confidence intervals and significance.

Appendix G.7 now reports 95% confidence intervals for all intrinsic dimensionality estimates and classification, regression, and forecasting scores. We use 1,000 bootstrap resamples, with dialogue-level resampling for MELD, DailyDialog, and IEMOCAP, and paired resampling for comparisons.

The results show that Continuous has higher intrinsic dimensionality than State in all 16 encoder–dataset settings, with non-overlapping confidence intervals; State has the lowest dimensionality in 14/16 settings. For the tuned nonlinearity probes, confidence intervals include zero in 32/48 cells, with positive and negative gaps in eight cells each (maximum absolute gap: 0.10). In the revised Table 3 (formerly Table 2), the winner's confidence interval lies entirely above the runner up's in 16/36 seed-42 comparisons; the remaining comparisons are marked as ties. Across three BERT seeds (42, 256, 512), task-score standard deviations are at most 0.031 and ID standard deviations at most 0.49. The best-performing representation is consistent across seeds in 11/13 settings, while State has the lowest ID on every dataset across all three seeds (Appendix G.7).

Q5. Can the DailyDialog temporal effect be separated from discourse or lexical patterns?

We thank the reviewer for encouraging a closer examination of the factors underlying the observed temporal patterns. Our additional analyses support interpreting the observed effect as discourse-position drift rather than emotion-specific temporal organization. We fitted the drift direction on training dialogues and evaluated it on held-out dialogues.

• On DailyDialog, within-dialogue Spearman correlation with turn position is 0.36–0.47 across all 12 encoder–objective settings, compared with 0.01–0.08 for directions fitted on shuffled dialogues. This difference is significant (p=0.005p=0.005, the minimum attainable with 200 random directions).

• The correlation remains 0.34–0.44 using only neutral utterances and 0.23–0.37 after controlling for emotion and dialogue act. Alignment with valence is at chance level (|cos| <= 0.16); the 95th percentile for random directions in 128 dimensions is approximately 0.17).

• The effect is comparable across State, Continuous, and Dynamic (0.40, 0.44, and 0.40, respectively). It is also unlikely to reflect topic recurrence, since topic remains constant within a DailyDialog dialogue. The effect is weaker on MELD (0.02–0.15) and IEMOCAP (0.07–0.21).

We have therefore revised the interpretation in Section 4.5, removed claims of the "temporal organization of emotion" from the abstract and conclusion, and report the finding as positional drift with the additional controls. We also propose removing 'Temporal' from the title to better reflect the revised scope and interpretation of the work and would welcome the Area Chair's view on this change.

We sincerely thank the reviewer for raising these important questions and engaging thoroughly with our work. The additional baselines, controls, and statistical analyses have enabled us to establish a more precise interpretation of our findings: training induces objective-dependent representational compression across four architectures; compactness is associated with categorical supervision rather than emotion alone; and emotional information remains target-specific even when representations have comparable dimensionality. We hope these revisions address the reviewer’s concerns and strengthen the empirical and interpretive contributions of the manuscript.

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

This paper studies how emotion is organized in the representation spaces of language models and how it evolves through conversation. Drawing on three theoretical perspectives on emotion (discrete categories, continuous dimensions, and conversational dynamics), the authors construct three representation objectives (State, Continuous, and Dynamic) on top of four encoders, all mapped to a common 128-dimensional latent space. These representations are analyzed on five affective datasets in terms of information accessibility, latent geometry, and temporal structure.

The authors report the following main findings. Emotional concepts are broadly accessible through linear decoding, and they concentrate in compact, structured subspaces. The resulting organization varies with the representation objective, while core geometric patterns are preserved across architectures. On this basis, the authors argue that the paper establishes a unified framework for characterizing emotion, linking the geometry of emotional states to representation learning and to their evolution over time within conversational context.

Critical strengths

The three-view modeling is complementary and meaningful. The paper operationalizes three theoretical perspectives on emotion (discrete categories, continuous dimensions, and conversational dynamics) as the State, Continuous, and Dynamic representation objectives. This allows the model’s behavior at different levels of emotion representation to be examined from different angles and compared against one another, which provides ample ground for discussion. The comparative experimental design is relatively comprehensive. The study covers 3 objectives, 4 encoders, and 5 datasets, and its evaluation spans downstream tasks, a probe ladder, geometric measures, and temporal analyses.

Critical weaknesses
  1. The methodological motivation is not sufficiently clear. The definitions of the three core representations (State, Continuous, and Dynamic) appear to be borrowed from external datasets and from long-standing conventions in emotion research. The paper seems to aim at exploring how accessibility and geometric structure differ across these three representation definitions. However, it remains unclear how the resulting conclusions relate to the real-world meaning of the three perspectives on emotion. The authors may wish to discuss this connection further.
  2. The analyses rely on dataset labels, which may bias the representation-level conclusions. The notion of “emotion” studied in this paper embeds several different views of emotion at once. Moreover, the datasets used do not uniformly support the three representation objectives: VAD annotations are available only in EmoBank and IEMOCAP; GoEmotions and EmoBank lack conversational context; and IEMOCAP is used only for external evaluation and has no training split. Nevertheless, the paper still reports results for all three representations on these datasets (e.g., Dynamic on GoEmotions, and State and Dynamic on EmoBank), without specifying on which data and with which labels each representation objective is trained.
  3. Some experiments do not fully support the corresponding core claims, for example the claim of “compact, structured subspaces”. The authors appear to use the observation that geodesic kNN classification outperforms Euclidean kNN as evidence that emotional representations have a low-dimensional structured organization. However, computing Δgeo (geodesic minus Euclidean kNN F1) from Table 6 shows that it is positive on DailyDialog and IEMOCAP but negative on GoEmotions and MELD, consistently across all four backbones. For BERT State, for example, the values are −0.076 (GoEmotions), −0.065 (MELD), +0.072 (DailyDialog), and +0.058 (IEMOCAP).
Questions for authors
  1. RBF-SVM in Fig.5 lies consistently the weakest probe. Could the authors explain this? A properly tuned RBF-SVM should generally be no weaker than a linear model, which suggests potential issues with the hyperparameter settings or with training-set subsampling.
  2. The purpose of Fig. 7 is not sufficiently clear. First, the three representations differ in their density patterns (State is compact, Continuous is smooth and localized, Dynamic is extended). This is to be expected, since the three use different training objectives, so differences in shape are almost inevitable; the figure therefore does not show that these differences are related to emotion. Second, the figure does not indicate emotion categories, so readers cannot judge whether these patterns reflect the organization of emotion. For example, does the “smoothness” of Continuous correspond to the valence–arousal structure predicted by dimensional theories of emotion? Does the “extension” of Dynamic correspond to emotional evolution in conversation?
  3. Figure quality. For example, like Figure 2 should be provided as a vector graphic.
  4. The intrinsic dimensionality estimator is not described. Please specify the estimator used.
5 comments
AuthorsOpenReview ↗

Responese to Weakness(W) 1 and 2

We sincerely thank the reviewer for their thorough evaluation of our work and for the valuable suggestions provided. The comments helped us identify several issues requiring clarification and, importantly, prompted additional analyses that have strengthened the paper's methodology and refined its conclusions. We revisited the experimental pipeline, verified that all 108 reported table entries reproduce the original results (all 108 table cells, max difference < 1e-7), and conducted additional controlled experiments. We respond to the points raised below. All corresponding revisions have been incorporated into the revised manuscript, with a version highlighting the changes provided in the supplementary material.

W1. How do the conclusions relate to the three theories of emotion?

We thank the reviewer for highlighting the need to clarify how our three representation objectives relate to established perspectives on emotion. Our work provides a systematic comparison of how representing emotion as a discrete category (State), continuous affective dimensions (Continuous), or a context dependent conversational state (Dynamic) shapes the structure, accessibility, and functional properties of learned neural representations.

Our additional analyses reveal distinct consequences of these objectives across four encoders and five datasets. State produces the most compact representations, with a mean intrinsic dimensionality (ID) of 6.4, compared with 11.2 for Continuous and 8.1 for Dynamic. Continuous retains higher dimensionality than State in all 16 encoder dataset settings, with non-overlapping 95% confidence intervals. It also exhibits lower local linearity (BERT: 0.43–0.49 versus 0.61–0.64 for State) and provides the strongest performance on VAD regression and held-out IEMOCAP transfer across all four encoders.

Dynamic, in contrast, does not exhibit evidence of emotion specific temporal organization beyond general conversational structure. The observed temporal regularity is more closely associated with dialogue position than emotional change, as discussed further in our response to Reviewer q1S5 (Q5).

These findings reveal distinct structural biases induced by the three objectives that are: categorical supervision favors compact representations, dimensional supervision preserves richer graded structure, and contextual supervision does not necessarily produce emotion specific temporal dynamics. We have incorporated this interpretation into the revised Section 4 and Conclusion.

W2. Which data and labels train each objective?

We thank the reviewer for highlighting the need to distinguish objective specific training from transfer evaluation. We have clarified this in Section 3 and added a summary of the training data and supervision. State is trained on GoEmotions (28 emotion labels; binary cross entropy), Continuous on EmoBank (valence, arousal, dominance; mean squared error), and Dynamic on MELD dialogue sequences (7 emotions; cross entropy over a GRU).

All other objective dataset combinations are transfer evaluations: the trained model is frozen, and a fixed linear probe is fitted on the target dataset (logistic regression, (C=1), or ridge regression, (alpha=1). Thus, Dynamic evaluated on GoEmotions uses MELD trained representations to decode GoEmotions labels.

The evaluated latent is the 128 dimensional projection of an individual utterance. The GRU provides contextual supervision during Dynamic training but is not used to compute the evaluated per-utterance representation. Equation 5 has been revised accordingly. VAD evaluation is restricted to EmoBank and IEMOCAP, while temporal analyses use conversational datasets. Since IEMOCAP has no official split, we use stratified five-fold cross-validation, excluding 2,510 turns labelled xxx'' or oth'' and the two ``disgust'' instances. The revised manuscript now specifies these training, transfer, and evaluation protocols.

AuthorsOpenReview ↗

Responese for Weakness(W) 3 and Question(Q) 1 & 2

W3. Δgeo changes sign across datasets.

We sincerely thank the reviewer for drawing attention to the inconsistent geodesic kNN results. On re-examining the implementation, we identified an error in the original geodesic probe. Each test point was connected to its nearest training point, which was then used as the prediction anchor. Since the anchor is itself the closest training node in graph distance, the procedure effectively reduced to Euclidean 1-NN. Consequently, predictions were identical in all 48 settings. The originally reported Δgeo therefore compared 1-NN against 15-NN rather than isolating the effect of geodesic distance.

We corrected the implementation to use a 15-NN vote based on kNN-graph geodesic distances and compared it against Euclidean 15-NN, thereby matching the neighbourhood size. Under this corrected comparison, the dataset-level mean differences (geodesic minus Euclidean macro-F1) are: GoEmotions, (-0.020); MELD, (-0.015); DailyDialog, (-0.009); and IEMOCAP, (-0.016). None of the 12 backbone–dataset settings per dataset have a 95% confidence interval entirely above zero. Across all 48 settings, geodesic kNN performs significantly worse in 25.

The correction removes the previously observed dataset dependent sign pattern, showing that the earlier differences were attributable to mismatched neighborhood sizes rather than a consistent advantage of geodesic distance. Accordingly, we have withdrawn the claim that geodesic neighborhoods improve classification, removed the corresponding column from the main results table, and retained the corrected analysis as a negative result in Appendix G.1. Importantly, our conclusions about representation compactness do not depend on geodesic kNN. They are supported independently by the intrinsic-dimensionality comparisons, their confidence intervals, and the pretrained-encoder baselines discussed further in our response to Reviewer q1S5 (W1).

Q1. Why is RBF-SVM the weakest probe?

We thank the reviewer for identifying that our original RBF-SVM comparison used fixed rather than tuned hyperparameters (C = 1, gamma = "scale"). We agree that this could disadvantage the nonlinear probe. We therefore repeated the comparison across all 48 settings using three-fold inner cross-validation, tuning C in {0.1, ..., 10^4} and gamma in {0.1, ..., 10} x (1/d). We applied the same tuning procedure to the linear probe's C, using the original training subsamples. The revised Figure 5 shows three important findings: • Tuning substantially improves RBF-SVM. Its mean macro-F1 increases by (+0.094), making it the strongest nonlinear probe in 41/48 settings.

• The overall conclusion remains unchanged. With both probes tuned, RBF-SVM exceeds the linear probe by less than 0.01 macro-F1 on average for every encoder. The mean nonlinearity gap is (+0.006). Its 95% CI includes zero in 32/48 settings and lies above zero in eight, six of which involve IEMOCAP. The largest positive gap is (+0.10) for MiniLM Dynamic.

• The selected hyperparameters provide further context. Tuning frequently favors smaller gamma and larger C, producing broader kernels that approach linear decision boundaries.

Importantly, the originally reported nonlinearity gap was determined by the best-performing nonlinear probe, which was an MLP in 47/48 settings, rather than the untuned RBF-SVM. Thus, tuning improves the fairness of the comparison without materially changing the central finding. Emotional information is broadly accessible through linear decoding, with modest nonlinear advantages concentrated in a minority of settings, predominantly IEMOCAP. We have included this interpretation in Section 4.1 and updated Figure 5.

Q2. Purpose of Fig. 7.

We thank the reviewer pointing out that the original Figure 7 did not sufficiently connect its visual patterns to emotional structure. We have redesigned it (now Figure 6 in the revised manuscript) to include emotion-colored points, class centroids, a complementary turn position visualization, and quantitative statistics for interpreting each panel. We also revisited the underlying analyses to address the two questions directly.

AuthorsOpenReview ↗

Responses to Question(Q) 3 and 4

Q3. Figure quality.

Fig. 2 has been re-exported at higher resolution to improve visual clarity. Should further refinement be necessary, we are happy to redraw the figure entirely in vector format for the camera-ready version.

Q4. Which intrinsic-dimension estimator?

We thank the reviewer for highlighting the need for greater methodological clarity. We use the TwoNN estimator (Facco et al., 2017), now explicitly identified in Section 3.3 and described further in Appendix A.

We have also clarified two preprocessing choices relevant to our datasets. First, exact duplicate utterances are removed before estimation because zero nearest neighbor distances can destabilize the estimator. Duplicates account for 18.2% of DailyDialog and 10.6% of MELD utterances. Second, we discard the upper 10% of distance ratios to reduce sensitivity to extreme values. Estimates use up to 3,000 training points, with 95% confidence intervals obtained from 20 subsamples.

For example, for BERT on GoEmotions, the estimated intrinsic dimensionalities are 7.15 [6.90, 7.39] for State and 15.76 [15.08, 16.14] for Continuous. These details and uncertainty estimates are now reported in the revised manuscript.

We are grateful to the reviewer for their insightful comments and the depth of engagement with our work. The comments led us to correct the geodesic analysis, establish a fairer tuned probe comparison, clarify the interpretation of our latent space visualizations, and document the intrinsic dimensionality estimation procedure more fully. These revisions have also helped us distinguish more carefully between general structural properties of learned representations and properties specifically associated with emotional information. We hope the revised analyses and clarifications adequately address the concerns raised by the reviewer. We would be pleased to provide any further clarification.

Comments on the paper

1 comment

Dear reviewers,

Thanks for your service! The authors have provided some comments now. Please take a look at them and assess to what extent they address your concerns. As a reminder, it is within your purview to decide whether score changes are warranted and whether you want to pose any follow-up questions.

Thanks!

— Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.