acceptodds
Under review as a conference paper at ICLR 2027

ACC-ERG:Coordinating Affective and Cognitive Support for Empathetic Response Generation

Abstract

Empathetic response generation methods model affective and cognitive information to better understand users' states. However, generated responses may mismatch the context by overlooking emotional needs or offering generic comfort when advice is required. To address this issue, we propose ACC-ERG, an Affective–Cognitive Coordination framework that adjusts the relative emphasis on affective and cognitive support according to the dialogue context. First, we build affective and cognitive experts to learn specialized support representations. Second, we design a Proportion Gate to predict a weight that characterizes the relative emphasis on affective versus cognitive support. Finally, we fuse these representations to generate an empathetic response according to the weight. Moreover, we construct ACC-ERG dataset with weight annotations and responses aligned with weights. Experiments on EmpatheticDialogues and ACC-ERG datasets show our framework outperforms state-of-the-art methods and demonstrate that varying weight correspondingly shifts the relative emphasis of affective and cognitive support in generated responses.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Empathetic Response Generation

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

This paper studies empathetic response generation and argues that prior methods improve user-state understanding but do not specify how affective and cognitive support should be composed in a response. Its main contributions are:

  1. A continuous formulation of context-adaptive support coordination, explicitly modeling the response-level affective–cognitive support composition via a weight w ∈ (0,1).

  2. The ACC-ERG framework, comprising two frozen specialized experts (an affective expert trained on EmpatheticDialogues and a two-stage cognitive expert trained on KardiaBench), a Proportion Gate that predicts the support weight, and a Fusion Bridge that fuses the two expert representations under that weight for response generation.

  3. The ACC-ERG dataset (4,998 instances derived from EmpatheticDialogues and KardiaBench), with GPT-4o-mini-assisted, human-reviewed annotations of the support weight, support intensity, and speech act, plus weight-aligned reference responses.

  4. Experimental evidence that ACC-ERG improves BLEU/Distinct over strong baselines on both ACC-ERG and EmpatheticDialogues, predicts support weights well (MAE 0.108, Spearman 0.780), and that intervening on the weight controllably shifts the affective vs. cognitive emphasis of generated responses.

Critical strengths
  1. This paper proposes manually annotated affective and cognitive weights, providing room for further optimization in subsequent work.

  2. A framework for explicitly modeling these weights is proposed.

Critical weaknesses
  1. The experimental comparisons in this paper involve certain unfair experimental settings, which require further clarification and discussion.
  2. Regarding the manual annotation, the inter-annotator agreement and the related evaluation criteria were not reported.
Questions for authors
  1. When training the cognitive expert and the affective expert, the full datasets—larger than the core dataset ACC-ERG proposed in this paper—were used (the 36,629 complete ED samples mentioned in Appendix C.1, and the 19,533 dialogues mentioned in Yuan et al., 2026). However, when comparing with the baselines, only the core dataset ACC-ERG was used rather than directly using the two merged original datasets, which results in different amounts of data being exposed to the baselines.

  2. Generally speaking, gold emotion labels are unavailable at test time. During the training of the affective and cognitive experts, the inputs both incorporate the emotional state; however, at inference time, such emotion labels are inaccessible. Theoretically, this gives rise to a severe discrepancy between training and inference, yet this issue is not addressed or explained in this paper.

  3. The method proposed in this paper aims to explicitly model the affective and cognitive aspects; however, it does not provide a targeted evaluation of the generation results after ablating either of the two components (e.g., whether ablating the affective expert genuinely impairs the affective capability, especially the human evaluation).

  4. The pre-annotation employed GPT-4o-mini, a model more capable than GPT-3.5-turbo which was released in 2022. This raises a natural question: if compared directly against GPT-4o-mini, would the proposed method still retain an advantage, particularly under one-shot or few-shot settings?

  5. During the annotation of the dataset proposed in this paper, human annotators provided corrections; however, the paper reports neither agreement-related metrics nor a reasonably feasible scoring rubric(there are no per-level anchor definitions or worked examples), which undermines the reproducibility of this work.

8 comments
AuthorsOpenReview ↗

Response to Q1

Dear Reviewer fBmD,

We greatly appreciate your hard work during the review process. Thank you very much for your valuable questions. We will address your concerns one by one to clarify any confusion.

Q1: The use of additional training data for the two experts creates an unfair comparison with baselines trained only on the ACC-ERG dataset.

A1: We thank the reviewer for pointing out the difference in training data exposure between ACC-ERG and the baselines. In the original setting, the two experts were trained on the full EmpatheticDialogues and KardiaBench datasets, while some baselines were trained only on ACC-ERG. To address this concern, we conducted a new experiment with controlled training data.

(1) Data-controlled training setup.

We introduced ACC-ERG (Data-Controlled), in which all training stages are restricted to the ACC-ERG training split, without using extra datasets for additional expert training. To preserve functional specialization, the affective and cognitive experts are trained separately on samples with clear affective-support and cognitive-support tendencies, respectively. The Proportion Gate and Fusion Bridge are then trained using the same ACC-ERG training split. This setting removes the additional exposure to the original datasets and provides a more comparable evaluation against baselines trained on ACC-ERG.

(2) Experimental results.

As shown in Table A1, ACC-ERG (Data-Controlled) achieves BLEU-1/2/3/4 scores of 47.51/33.80/26.45/21.38, remaining better than the main baselines reported in the original submission. Although some generation metrics and support-priority correlations decrease slightly, the data-controlled variant maintains strong overall performance. These results suggest that additional experts training data are beneficial but do not fully account for the performance gains of ACC-ERG.

Table A1: Performance comparison with baselines.

ModelsB-1↑B-2↑B-3↑B-4↑D-1↑D-2↑
CEM30.0518.3413.029.122.889.59
CASE31.7819.5114.0010.692.488.03
IAMM29.2816.8711.388.335.7517.83
ReDiff31.5519.8114.5511.110.651.67
GPT-3.5-Turbo40.5625.4918.8113.449.3541.21
ACC-ERG (Data-Controlled)47.5133.8026.4521.389.1935.20
ACC-ERG (Original)48.1434.3827.0221.869.0334.71

We will include the data-controlled experiment in the revised manuscript and clearly distinguish the training settings of the original ACC-ERG and ACC-ERG (Data-Controlled). We will also report their performance under both settings to make the comparisons and the contribution of additional training data more transparent.

AuthorsOpenReview ↗

Response to Q3

Q3: The paper lacks targeted human evaluation to verify whether removing each expert affects its corresponding support capability.

A3: We thank the reviewer for this suggestion. To address this concern, we conducted an additional human evaluation of the expert ablations.

(1) Human evaluation setup.

We selected 200 test samples and recruited five independent evaluators to conduct pairwise comparisons between the full ACC-ERG model and variants without the affective or cognitive expert. The generated responses were evaluated on three aspects: Affective Support (Aff.), Cognitive Support (Cog.), and Appropriateness (App.). We report the win, tie, and loss rates of the full model against each ablated variant.

(2) Results and analysis.

As shown in Table A3, compared with the model without the affective expert, ACC-ERG achieves a 62.5% win rate in affective support, compared with 45.2% in cognitive support. Conversely, when the cognitive expert is removed, the full model achieves a 64.3% win rate in cognitive support, compared with 44.1% in affective support.

The full model also achieves appropriateness win rates of 59.6% and 60.8% against the variants without the affective and cognitive experts, respectively.

These results indicate that removing the affective expert has a greater impact on affective support, while removing the cognitive expert has a greater impact on cognitive support. This pattern is consistent with the intended functional specialization of the two experts and shows that both contribute to response appropriateness.

Table A3: Human evaluation of expert ablations. Win/Tie/Lose indicate the preference rates of the full ACC-ERG model over the corresponding ablated variant.

ComparisonAspectWin (%)Tie (%)Lose (%)
vs. w/o Affective ExpertAff.62.524.213.3
Cog.45.238.316.5
App.59.626.513.9
vs. w/o Cognitive ExpertAff.44.140.015.9
Cog.64.321.614.1
App.60.824.215.0

We will include these human evaluation results and the corresponding analysis in the revised manuscript to provide more direct evidence of the experts' differentiated contributions.

AuthorsOpenReview ↗

Response to Q2

Q2: The use of gold emotion labels during expert training may introduce a discrepancy between training and inference, and the paper does not clarify how emotions are inferred at test time.

A2: We thank the reviewer for raising this concern. We agree that gold emotion labels are generally unavailable in real-world empathetic response generation. We therefore clarify how emotional information is modeled and used at different stages of ACC-ERG.

(1) Emotion understanding at test time.

At test time, ACC-ERG does not use a separate emotion classifier or require gold emotion labels. As described in Section 3 and illustrated in Figure 2, the model directly takes the situation, dialogue history, and current user utterance as input. The affective and cognitive experts encode the dialogue context and extract emotion-related representations using the parameters learned through pretraining and fine-tuning. These representations are then used for support-priority prediction and empathetic response generation.

Thus, ACC-ERG models emotional information implicitly through contextual representations, rather than relying on externally provided emotion labels at test time.

(2) Distinction between emotion understanding and support coordination.

We would like to emphasize that our primary goal is not to explicitly predict emotion categories, but to coordinate affective and cognitive support in generated responses based on the user's emotional state and situation.

During expert training, emotion labels are used as additional contextual information to help the experts learn relationships between emotional states, situations, intentions, and potential causes. Specifically, the affective expert focuses on emotion understanding and affective support, while the cognitive expert focuses on situational understanding, reasoning about potential causes, and cognitive support.

During response generation, the Proportion Gate predicts the relative priority of affective and cognitive support from the dialogue context, and the Fusion Bridge coordinates the two expert representations accordingly. The central objective is to adapt the support emphasis to the user's current needs, rather than treating emotion classification as a prerequisite for response generation.

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

This paper attempt to address the mismatch between users’ emotional needs and the types of support provided in empathetic response generation by proposing an affective-cognitive coordination framework and a new dataset with support-proportion annotations.

Critical strengths

The paper attempts to explicitly model support preferences. The proposed Proportion Gate aims to adaptively balance affective and cognitive support according to dialogue context, offering a mechanism for controlling the emphasis of different types of support. The construction of the ACC-ERG dataset, with support-proportion annotations and corresponding responses, 'may' also provide a useful resource for investigating this problem setting.

Critical weaknesses
  1. The paper introduces affective and cognitive support as two dimensions of empathetic response generation. However, the theoretical motivation for this decomposition remains insufficiently developed. The introduction primarily cites two prior studies (Line 33) and reports two relevant cases (Lines 044–046), which does not adequately establish why these two dimensions constitute an appropriate or sufficiently comprehensive representation of empathetic support. The authors are encouraged to explore the correlations of the two aspectes from the view of their own perspectives.

In particular, the authors should provide stronger theoretical grounding, potentially drawing on established psychological theories of emotional support and coping, to justify this decomposition. Without such justification, the necessity and significance of explicitly coordinating these two dimensions remain unclear.

  1. As illustrated in Fig. 1, the framework represents affective and cognitive support as two components whose relative weights sum to one. A simple gating mechanism that assigns relative weights to their representations may be insufficient to capture their potentially complex interactions. From a methodological perspective, the framework appears to consist primarily of two representation-learning branches followed by weighted fusion.

  2. The authors state that 'existing datasets do not provide direct supervision for response-level support proportions and therefore construct the ACC-ERG dataset with proportion annotations and corresponding responses'. However, this rationale raises a fundamental concern: the proposed framework and dataset appear to be built around a newly introduced problem formulation whose theoretical and practical necessity has not been sufficiently established.

  3. As illustrated in Fig. 2, the two experts appear to have largely parallel architectures. However, the paper does not sufficiently explain how these modules are encouraged to learn distinct and specialized capabilities corresponding to affective and cognitive support. Merely assigning different names to two representation-learning branches does not establish that they function as genuine experts.

Questions for authors

see 'Weaknesses'.

4 comments
AuthorsOpenReview ↗

Response to W1

Dear Reviewer VGkd,

We greatly appreciate your hard work during the review process. Thank you very much for your valuable suggestions. We will address your concerns one by one to clarify any confusion.

Q1: The theoretical motivation for decomposing empathetic support into affective and cognitive dimensions, and the necessity of explicitly coordinating them, remain insufficiently justified.

A1: We thank the reviewer for this suggestion. We will strengthen the theoretical motivation by drawing on established research on empathy, social support, coping, and supportive communication.

(1) Theoretical basis for affective and cognitive support.

Davis (1983) and Decety and Jackson (2004) describe affective and cognitive empathy as distinct but interacting components of human empathy. This distinction provides the psychological basis for our approach.

From the perspective of social support theory, House (1981) distinguishes emotional, informational, and appraisal support. We broadly group emotional validation and care under affective support, and explanation, perspective taking, and cognitive reframing under cognitive support. Stress and coping theory (Lazarus & Folkman, 1984) further distinguishes emotion-focused and problem-focused coping, suggesting that individuals may need help with both emotional distress and understanding or addressing stressful situations.

Moreover, supportive communication research (Burleson & Goldsmith, 1998; Jones & Wirtz, 2006) shows how emotional comforting and cognitive reappraisal are related in the process of emotional adjustment. These findings support our view that affective and cognitive support serve different but complementary functions.

Thus, we consider affective and cognitive support as two broad functional dimensions of empathetic responses. Affective support primarily addresses emotional experiences, while cognitive support focuses on understanding and interpreting situations.

(2) Why explicitly coordinate the two dimensions?

The optimal matching perspective (Cutrona & Russell, 1990) suggests that effective support should match the recipient's needs and circumstances. Even when users express similar emotions, their support needs may differ. For example, one user may seek emotional validation, while another may need help understanding or resolving a problem.

This motivates us to model the relative emphasis on affective and cognitive support as a context-adaptive weight. Accordingly, ACC-ERG learns to predict this priority from dialogue context and uses it to coordinate the two support representations. We further construct the ACC-ERG dataset with support-priority annotations and need-aligned responses to provide supervision for this process.

We will clarify these theoretical motivations and the scope of our functional decomposition in the revised manuscript.

References

[1] Davis, M. H. (1983). Measuring individual differences in empathy: Evidence for a multidimensional approach. Journal of Personality and Social Psychology, 44(1), 113–126.

[2] Decety, J., & Jackson, P. L. (2004). The functional architecture of human empathy. Behavioral and Cognitive Neuroscience Reviews, 3(2), 71–100.

[3] House, J. S. (1981). Work Stress and Social Support. Addison-Wesley.

[4] Lazarus, R. S., & Folkman, S. (1984). Stress, Appraisal, and Coping. Springer Publishing Company.

[5]Burleson, B. R., & Goldsmith, D. J. (1998). How the comforting process works: Alleviating emotional distress through conversationally induced reappraisals. In P. A. Andersen & L. K. Guerrero (Eds.), Handbook of Communication and Emotion: Research, Theory, Applications, and Contexts (pp. 245–280). Academic Press.

[6] Jones, S. M., & Wirtz, J. G. (2006). How does the comforting process work? An empirical test of an appraisal-based model of comforting. Human Communication Research, 32(3), 217–243.

[7] Cutrona, C. E., & Russell, D. W. (1990). Type of social support and specific stress: Toward a theory of optimal matching. In Social Support: An Interactional View (pp. 319–366). Wiley.

AuthorsOpenReview ↗

Response to W2

Q2: The proposed framework appears to rely mainly on weighted fusion, which may be insufficient to capture complex interactions between affective and cognitive support.

A2: We agree that affective and cognitive support serve different but complementary functions, and modeling their interactions is important. This is one of the main motivations for introducing the Fusion Bridge.

(1) The fusion mechanism is not limited to weighted averaging.

As described in Section 4.3, our Fusion Bridge combines Explicit Linear Routing with nonlinear Residual Interaction. The Proportion Gate first predicts the relative priority of affective and cognitive support from the dialogue context and uses it to construct a weighted representation. The Fusion Bridge then takes the two expert representations, their difference, their element-wise product, and the predicted support weight as inputs to a nonlinear residual network. In addition, a priority-conditioned scaling network adjusts the contribution of the residual interaction according to the predicted weight.

This design preserves the interpretability of support-priority weighting while allowing the model to learn cross-branch interactions beyond simple linear fusion.

(2) Evidence from ablation experiments.

As shown in Table 1, ACC-ERG achieves a BLEU-4 score of 21.86. Removing the Fusion Bridge while retaining support-priority prediction and linear routing reduces BLEU-4 to 20.37. Removing the Proportion Gate also decreases BLEU-4 to 20.84. These results indicate that both priority-based routing and nonlinear interaction contribute to generation performance, with the Fusion Bridge providing a larger improvement in this evaluation.

Overall, ACC-ERG combines interpretable priority-based routing with nonlinear cross-branch interaction. The former controls the relative emphasis on affective and cognitive support, while the latter captures complementary information and interactions between the two expert representations.

AuthorsOpenReview ↗

Response to W3

Q3: The theoretical and practical necessity of explicitly modeling support proportions and constructing the ACC-ERG dataset remains unclear.

A3: We thank the reviewer for raising this concern. We have discussed the theoretical basis for affective and cognitive support in our response to W1. Here, we further clarify the motivation for explicit support weight modeling and the role of the ACC-ERG dataset.

(1) Motivation from mismatched support in empathetic responses.

Existing methods have improved the understanding of users' emotions and situations by modeling emotional states, emotion causes, commonsense knowledge, and their interactions. However, understanding a user's state does not necessarily lead to a response that matches their support needs. For example, a model may provide excessive reassurance when the user seeks practical guidance, or offer explanations and advice too early when the user primarily needs emotional validation. These cases highlight the importance of not only understanding the user but also determining the appropriate emphasis of different support functions in the response.

(2) Why explicitly model support priority?

The optimal matching perspective (Cutrona & Russell, 1990) suggests that effective support should match the recipient's needs and circumstances. Motivated by this view, we represent the relative emphasis on affective and cognitive support as a continuous, context-dependent support priority. This allows the model to explicitly learn how the appropriate balance between the two support functions varies across dialogue contexts, rather than relying solely on implicit representation learning.

(3) Why construct the ACC-ERG dataset?

Existing empathetic dialogue datasets mainly provide dialogue contexts and reference responses. Although they contain rich emotional and contextual information, they generally lack explicit annotations of response level affective–cognitive support emphasis. We therefore construct the ACC-ERG dataset with support weight annotations and corresponding need aligned responses. These annotations provide direct supervision for weight prediction and empathetic response generation, while also enabling us to evaluate whether the model can adjust its support emphasis according to different user needs.

In summary, our goal is not to introduce a new proportion parameter for its own sake, but to address the mismatch between understanding users' states and providing appropriate support. The continuous support weight offers an interpretable way to model this problem, while the ACC-ERG dataset provides the supervision needed to study it. We will clarify this motivation and the role of the dataset in the revised manuscript.

AuthorsOpenReview ↗

Response to W4

Q4: The paper does not sufficiently explain how the two structurally similar branches learn distinct affective and cognitive support capabilities.

A4: We thank the reviewer for this comment. Although the two experts share the same backbone architecture, their functional specialization is encouraged through different training data and supervision objectives. We further clarify this design and provide additional human evaluation results.

(1) Functional specialization through different training objectives.

As described in Section 4.1, the affective expert is fine-tuned on EmpatheticDialogues to learn emotion recognition, emotional validation, and supportive expression. The cognitive expert is trained on KardiaBench in two stages: first, it learns to understand user intentions, contextual information, potential causes, and support directions through structured cognitive analysis; then, it learns to generate natural-language responses using this cognitive supervision. The two experts have separate LoRA parameters, which remain frozen during subsequent coordination training. Thus, their specialization is encouraged by different supervision objectives rather than architectural differences.

(2) Evidence from additional human ablation evaluation.

We conducted an additional human evaluation on 200 test samples with five independent evaluators. The full ACC-ERG model was compared against variants without either expert on three aspects: Affective Support (Aff.), Cognitive Support (Cog.), and Appropriateness (App.).

As shown in Table A4, compared with the model without the affective expert, ACC-ERG achieves a 62.5% win rate in affective support, compared with 45.2% in cognitive support. Conversely, when the cognitive expert is removed, ACC-ERG achieves a 64.3% win rate in cognitive support, compared with 44.1% in affective support. The full model also achieves appropriateness win rates of 59.6% and 60.8% against the two ablated variants, respectively.

These results show that removing each expert has a greater impact on its corresponding support dimension. This provides direct evidence of their differentiated functional contributions, rather than simply assigning different names to otherwise identical branches.

(3) Evidence from support-weight intervention.

Figure 4 provides further evidence that the model can distinguish and coordinate the two support functions. As the affective support weight increases from 0.1 to 0.9, the affective support score rises from 3.15 to 3.91, while the cognitive support score decreases from 3.87 to 2.98. Contextual relevance and fluency remain relatively stable. These results show that adjusting the support weight can shift the emphasis of generated responses between affective and cognitive support.

Overall, the two experts are encouraged to learn different capabilities through specialized supervision. The additional human ablation results demonstrate their differentiated contributions, while the weight intervention results show that ACC-ERG can coordinate these representations to control response emphasis. We will clarify the specialization mechanism and include the additional evaluation results in the revised manuscript.

Table A4: Human evaluation of expert specialization. Win/Tie/Lose indicate the preference rates of the full ACC-ERG model over the corresponding ablated variant.

ComparisonAspectWin (%)Tie (%)Lose (%)
vs. w/o Affective ExpertAff.62.524.213.3
Cog.45.238.316.5
App.59.626.513.9
vs. w/o Cognitive ExpertAff.44.140.015.9
Cog.64.321.614.1
App.60.824.215.0
Reviewer_iZZGOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents ACC-ERG, a framework for coordinating affective and cognitive support in empathetic response generation. It combines two specialized experts through a gating module that predicts a context-dependent support weight, which guides the fusion of their representations for response generation. The authors also construct a dataset with support-weight annotations and corresponding responses, and assess generation quality and controllability through automatic evaluation, human assessment, and weight-intervention experiments.

Critical strengths
  1. The research problem has practical relevance. The paper investigates how empathetic responses can align their support with users’ current needs, addressing issues such as premature advice and generic reassurance. Coordinating affective and cognitive support is a meaningful direction with clear application value.

  2. The framework provides an explicit mechanism for controlling support emphasis. An explicit weight guides the fusion of the two expert representations, allowing the relative emphasis on affective and cognitive support to be adjusted directly. The weight-intervention experiments provide preliminary evidence of this controllability.

  3. Human evaluation complements the automatic metrics. The paper conducts anonymous A/B comparisons across empathy, contextual relevance, fluency, and support appropriateness. These assessments provide preliminary evidence of the full model’s overall response quality relative to the selected baselines.

Critical weaknesses
  1. The reliability of the support-weight annotations is insufficiently established. Continuous support weights provide a central supervision signal, yet the paper does not report inter-annotator agreement, the proportion of model-generated annotations revised by human reviewers, or whether adjacent weight levels can be meaningfully distinguished. The results demonstrate that the model can predict the provided labels, but do not establish that these labels consistently and reliably represent appropriate support compositions.
  2. The lack of controlled baselines makes the source of the performance gains unclear. ACC-ERG combines Flan-T5-XL with expert training on different data sources and teacher-generated response targets. However, the evaluation lacks direct fine-tuning and simple weight-conditioned generation baselines using the same backbone and training data. Consequently, the current comparisons do not adequately isolate the contribution of the proposed coordination mechanism from the benefits of additional training supervision.
  3. The contribution of adaptive coordination to support appropriateness has not been directly established. BLEU primarily measures lexical overlap with reference responses and cannot directly determine whether the support provided matches the user’s needs. Although the paper includes human comparisons against baselines, weight-intervention analyses, and fixed-routing experiments, it does not provide a human evaluation of support appropriateness under predicted, fixed, and shuffled weights. The existing evidence therefore does not directly establish whether adaptive weighting produces responses that better address users’ current support needs.
Questions for authors

Could you provide inter-annotator agreement statistics for the support weights, the proportion of LLM-generated labels revised by human reviewers, and evidence that annotators can reliably distinguish adjacent weight levels?

How does ACC-ERG compare with a single Flan-T5-XL model directly fine-tuned on the same training data and response targets, with and without explicit weight conditioning? Such comparisons would help isolate the benefits of the coordination mechanism from those of additional supervision.

Beyond the reported automatic fixed-routing results, is there human evaluation evidence that predicted weights yield more appropriate support than fixed or randomly shuffled weights for the same dialogue contexts?

4 comments
AuthorsOpenReview ↗

Response to W1&Q1

Dear Reviewer iZZG,

We greatly appreciate your hard work during the review process. We greatly value your questions and appreciate the opportunity to discuss our work with you.We will address your concerns one by one to clarify any confusion.

Q1: The reliability of support weight annotations remains unclear, particularly regarding inter-annotator agreement, human revisions of LLM-generated labels, and the distinction between adjacent weight levels.

A1: We thank the reviewer for raising this concern. Since support-weight annotations provide the main supervision for ACC-ERG, we further clarify the annotation process and provide additional evidence of annotation consistency.

(1) Annotation guideline development and calibration.

We define the support weight w∈{0.1,0.2,…,0.9}w\in\{0.1,0.2,\ldots,0.9\} as the relative emphasis on affective support in an ideal response. Two annotators independently labeled two non-overlapping sets of 50 stratified samples, totaling 100 instances. After each round, we reviewed disagreements and refined the guidelines, particularly the distinctions among emotional disclosure, validation seeking, advice seeking, and social sharing. We also clarified that emotional intensity does not directly determine affective-support priority.

In the second round, 36 of the 50 samples (72%) received annotations differing by no more than one weight level, with a mean absolute difference of 0.142. The first-round agreement rate and mean absolute difference are being rechecked against the original annotation records. We will report the verified results and provide the complete annotation guidelines in the appendix.

(2) LLM-assisted pre-annotation.

After finalizing the guidelines, we provided GPT-4o-mini with the annotation instructions and eight human-adjudicated examples covering different support priorities. With temperature set to 0, the model generated structured pre-annotations for 5,000 candidate instances, including affective and cognitive support priorities, speech acts, and brief rationales. The finalized annotations were then used to construct need-aligned responses.

(3) Human review and revisions.

Two reviewers independently examined all 5,000 pre-annotations together with their dialogue contexts. Among these instances, 4,992 (99.84%) required no revision, 6 (0.12%) were flagged for disagreements about support priority, and 2 (0.04%) were excluded because of inconsistencies between the situation and dialogue history. The six disputed cases were independently adjudicated by a third reviewer, who determined the final labels. After excluding the two invalid instances, the dataset contains 4,998 valid samples.

(4) Distinguishing adjacent weight levels.

To distinguish adjacent weights, the guidelines use two counterfactual questions: whether a response would be inadequate without emotional acknowledgment, and whether it would overlook the user's request without explanation or practical guidance.

For example, w=0.2w=0.2 indicates that guidance should dominate, with only brief emotional acknowledgment; w=0.3w=0.3 still prioritizes guidance, but emotional acknowledgment plays a more meaningful role; and w=0.4w=0.4 indicates that both support functions are needed, although cognitive support remains slightly more important. The same principle applies to affective-dominant weights.These criteria provide an interpretable distinction between neighboring weight levels, although fine-grained judgments may still vary between annotators.

We will include the annotation guidelines, representative examples, calibration results, and human review statistics in the revised manuscript to improve transparency and reproducibility.

AuthorsOpenReview ↗

Response to W2&Q2

Q2: The lack of controlled baselines makes the source of the performance gains unclear.

A2: To address this concern, we conducted additional experiments using a parameter-matched single LoRA baseline and a data controlled variant to further examine the sources of ACC-ERG's performance gains.

(1) Comparison with a parameter-matched single LoRA baseline.

We introduced a Single-LoRA baseline based on the same Flan-T5-XL backbone. To control for the effect of trainable parameter count, we set the LoRA rank to 52, making its number of trainable parameters comparable to that of ACC-ERG. The results show that Single-LoRA achieves BLEU-1/2/3/4 scores of 41.91, 28.32, 21.29, and 16.57, respectively, while ACC-ERG achieves 48.14, 34.38, 27.02, and 21.86. These results indicate that single-branch parameter-efficient fine-tuning with a comparable number of trainable parameters does not achieve the same generation performance as ACC-ERG. This suggests that the performance gains are not solely due to an increase in the number of trainable parameters.

(2) Comparison with the original Flan-T5-XL.

We further evaluated the original Flan-T5-XL without task-specific fine-tuning to examine its performance on empathetic response generation. Its scores are both substantially lower than those of Single-LoRA and ACC-ERG. This comparison provides a reference for the performance of the backbone without task-specific adaptation and highlights the importance of fine-tuning for empathetic response generation.

(3) Controlling for the effect of additional experts training data.

To investigate the potential contribution of the additional data used for expert training, we further constructed a data-controlled variant, ACC-ERG (Data-Controlled). All training stages of this variant are restricted to the ACC-ERG dataset. The two experts are trained separately using endpoint samples with clear affective-support and cognitive-support tendencies, without using the full EmpatheticDialogues and KardiaBench datasets for additional expert training. The results show that ACC-ERG (Data-Controlled) achieves a BLEU-4 score of 21.38, close to the 21.86 achieved by the original ACC-ERG, while remaining better than the main baselines in the original submission. This result suggests that ACC-ERG can maintain strong generation performance without additional expert-training data, and that its performance advantage does not solely depend on a larger training dataset.

Regarding the simple weight-conditioned generation baseline.

We also carefully considered the reviewer's suggestion of a simple weight-conditioned generation baseline. In our literature review, we have not identified directly comparable studies that predict continuous relative priorities between affective and cognitive support from dialogue context and use these predicted priorities to coordinate specialized expert representations.

These additional experiments and analyses provide a more controlled comparison of ACC-ERG with alternative settings and help clarify the respective effects of model capacity, additional training data, and the proposed support coordination mechanism.We will include the corresponding experimental results in the revised manuscript.

Table 1: Comparison with controlled baselines and component variants.

ModelB-1 ↑B-2 ↑B-3 ↑B-4 ↑D-1 ↑D-2 ↑
Flan-T5-XL21.2411.577.364.589.6935.10
Single-LoRA41.9128.3221.2916.578.8333.48
w/o Cognitive Branch47.1733.6926.3621.219.2835.35
w/o Affective Branch47.8633.9526.4621.229.2335.20
w/o Gate46.5033.0525.8320.848.8033.93
w/o Fusion Bridge47.0733.3325.7020.378.7034.30
ACC-ERG (Data-Controlled)47.5133.8026.4521.389.1935.20
ACC-ERG (Ours)48.1434.3827.0221.869.0334.71
AuthorsOpenReview ↗

Response to W3&Q3

Q3: The paper lacks direct human evaluation showing whether predicted support weights improve response appropriateness compared with fixed or shuffled weights.

A3: We thank the reviewer for this suggestion. So we conducted an additional human evaluation comparing ACC-ERG with predicted weights against variants using fixed weights.

(1) Human evaluation setup.

We sampled 200 dialogue contexts according to the support weight distribution of the training set, covering low (0.1–0.3), medium (0.4–0.6), and high (0.7–0.9) priority regions, with 35, 108, and 57 samples, respectively. Five independent evaluators compared responses generated using predicted weights against those generated using fixed weights of 0.2, 0.5, 0.529, and 0.8. The evaluation covered Empathy, Relevance, Fluency, and Appropriateness, with particular attention to whether the response matched the user's current support needs. In particular, the shuffled-weight setting is designed to examine whether preserving the correspondence between dialogue context and support weight is important for generating appropriate responses.

(2) Results on support appropriateness.

As shown in Table A3, predicted weight shows clear advantages over fixed weights when the fixed setting differs from the user's support needs.

For low weight contexts, ACC-ERG achieves a 71.4% win rate in appropriateness against the fixed weight of 0.8. For high weight contexts, it achieves a 66.7% win rate against the fixed weight of 0.2. In medium weight contexts, predicted weight also outperforms the fixed weight of 0.2, with a 57.4% win rate.

However, predicted weight does not consistently outperform fixed weights close to the corresponding weight region. For example, the fixed weight of 0.5 performs better in medium weight contexts, while the fixed weight of 0.8 performs better in high weight contexts.

These results suggest that the appropriate support emphasis depends on the dialogue context. A fixed weight may work well for a particular group of users but does not necessarily generalize across different needs. Predicted weight allows ACC-ERG to adjust its support emphasis according to the input rather than applying the same configuration to all responses.

Table A3: Human evaluation of support appropriateness across different priority regions. Win/Tie/Lose (%) compare ACC-ERG with predicted weights against fixed-weight variants.

Regionvs. Fixed 0.2vs. Fixed 0.5vs. Fixed 0.529vs. Fixed 0.8
Low (35)34.3 / 20.0 / 45.757.1 / 14.3 / 28.660.0 / 14.3 / 25.771.4 / 8.6 / 20.0
Medium (108)57.4 / 13.9 / 28.736.1 / 18.5 / 45.431.5 / 18.5 / 50.046.3 / 16.7 / 37.0
High (57)66.7 / 12.3 / 21.154.4 / 14.0 / 31.650.9 / 14.0 / 35.138.6 / 15.8 / 45.6

(3) Implications for adaptive coordination.

The results provide direct human-evaluation evidence that support appropriateness is sensitive to the choice of support weight. They also show why a single fixed weight is not suitable for all dialogue contexts. While predicted weight is not always the best choice within every weight region, it provides a context-adaptive mechanism for adjusting support emphasis without selecting a fixed weight separately for each group.

We will include these human evaluation results and the corresponding analysis in the revised manuscript to further clarify the contribution and limitations of adaptive support coordination.

Dear Reviewer iZZG,

We sincerely thank you once again for your valuable feedback. To address your concerns, we have provided further clarifications and additional experimental evidence. We truly appreciate the opportunity to engage in this discussion with you, and we would be grateful for any further feedback you may have. We look forward to hearing your thoughts.

Best regards,

Authors of Paper #20602

Comments on the paper

1 comment

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

We sincerely thank all reviewers for their valuable comments and feedback !

In this paper, we propose ACC-ERG, which explicitly models and coordinates the relative emphasis of affective and cognitive support in empathetic responses according to the dialogue context. The main contributions of our work are as follows:

  1. A response-level support coordination formulation: We address the gap between understanding users' states and matching their actual support needs in empathetic response generation, explicitly formulating context-adaptive coordination of affective and cognitive support as a learning problem.

  2. An interpretable and controllable support coordination framework: ACC-ERG integrates two functionally specialized experts, a Proportion Gate for predicting support weights, and a Fusion Bridge that combines explicit weight-guided routing with nonlinear residual interaction to coordinate affective and cognitive support representations.

  3. A dataset with support weight annotations and aligned responses: We construct the ACC-ERG dataset containing 4,998 valid dialogue instances, providing support weight annotations and need-aligned training responses to supervise context-dependent support coordination.

Following the reviewers' suggestions, we conducted the following additional experiments and expanded our analyses during the rebuttal period:

  1. Parameter-matched and data-controlled baselines: ACC-ERG outperforms the parameter-matched Single-LoRA baseline on all BLEU-1/2/3/4 and Distinct-1/2 metrics. Furthermore, ACC-ERG (Data-Controlled), which does not use the full EmpatheticDialogues and KardiaBench datasets for additional expert training, still achieves better automatic evaluation results than the other comparison baselines. These results demonstrate that the performance advantages of ACC-ERG are not solely attributable to increased trainable parameters or additional training data. (Section 5.1; Appendix D.2; Reviewer fBmD)

  2. Targeted human evaluation of expert specialization: We conducted pairwise human evaluations on 200 test samples with five independent evaluators. Compared with the variant without the affective expert, the full ACC-ERG model achieves a 62.5% win rate in affective support. Compared with the variant without the cognitive expert, it achieves a 64.3% win rate in cognitive support. The corresponding appropriateness win rates are 59.6% and 60.8%, respectively, further validating the differentiated functions of the two experts and their contributions to response quality. (Appendix D.1; Reviewer fBmD)

  3. Human evaluation of adaptive versus fixed support weights: Across dialogue contexts with low, medium, and high support weights, dynamic routing shows clear advantages when fixed weights do not match users' support needs. For example, in low-weight contexts, dynamic routing achieves a 71.4% win rate in response appropriateness against Fixed-0.8, while in high-weight contexts, it achieves a 66.7% win rate against Fixed-0.2. We also report cases where fixed weights closer to the corresponding weight regions perform better, further demonstrating that response appropriateness depends on the alignment between support weights and dialogue contexts. (Appendix D.3; Reviewer iZZG)

  4. Comparison with GPT-4o-mini under different prompting settings: We conducted additional experiments with GPT-4o-mini under zero-shot, one-shot, four-shot, and six-shot settings on the same 469 test samples. ACC-ERG outperforms GPT-4o-mini zero-shot on all BLEU metrics and maintains competitive generation performance against its few-shot settings without requiring in-context demonstrations at inference time. (Appendix D.5; Reviewer fBmD)

  5. Clarifications on annotation reliability and dataset construction: We refined the support-weight annotation guidelines, supplemented the human annotation calibration results, clarified the criteria for nine support-priority levels, and documented the LLM-assisted pre-annotation and human review procedures in greater detail to improve the transparency and reproducibility of dataset construction. (Appendix B; Reviewer iZZG, Reviewer fBmD)

  6. Further Clarifications on Dual-Expert Design: Following additional feedback from Reviewers fBmD and VGkd, we provided further clarifications on why two structurally similar encoders can learn distinct affective and cognitive support capabilities and how our dual-expert architecture is designed to model these two functions.

We have incorporated the additional experimental setups, results, and analyses into the revised manuscript, with all modifications highlighted in blue. We hope these additions adequately address the reviewers' concerns and further clarify the motivation, effectiveness, and controllability of ACC-ERG.

Best regards,

Authors of Paper #20602

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.