acceptodds
Under review as a conference paper at ICLR 2027

Understanding Distress Is Not Expressing Empathy: Supportive Empathic Response Neurons in LLMs

Abstract

Large language models (LLMs) can recognize user distress and generate empathic responses. However, whether these behaviors rely on shared or separable internal mechanisms remains unclear. Because input-encoded information is not necessarily utilized during generation, we distinguish distress representation from empathic response realization. To isolate this response-side mechanism, we introduce a counterfactual difference-in-differences (DiD) activation framework. This framework contrasts matched supportive and neutral responses across distress and non-distress contexts. In this way, it mitigates confounds from input-side sensitivity and generic supportive language. The procedure identifies a sparse population of post-gating feed-forward network (FFN) units, termed Supportive Empathic Response (SER) neurons. Experiments demonstrate that SER neurons preferentially track affective empathic expression rather than instrumental support or generic supportive content, while remaining largely invariant to distress magnitude. Crucially, direct activation interventions across all four model families demonstrate reproducible bidirectional causal effects on both expressed empathy and support-strategy realization: suppressing frozen SER neurons reduces EPITOME-based expressed empathy by 10.3%–31.9% relative to baseline and shifts ESConv responses from affective toward instrumental support, whereas amplification consistently produces the opposite pattern. Furthermore, these interventions have only limited effects on general reasoning and broader emotion-related capabilities. Together, these findings identify a sparse, causally consequential response-side mechanism and provide evidence that representing user distress and realizing an empathic response are supported by partially dissociable internal mechanisms in LLMs.

Primary area:
interpretability and explainable AI
Keywords:
Neuron, Empathy, Causal Intervention, Interpretability, Large Language Models

Reviews

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

The paper investigates internal components involved in empathic response generation and their relationship to distress representation. It identifies Supportive Empathic Response (SER) neurons using a difference-in-differences contrast between matched supportive and neutral responses in distress and non-distress contexts. The selected FFN populations are frozen before held-out and downstream evaluation. Across four model families, independent empathy and support-strategy datasets, a factorial response experiment, and activation interventions provide evidence that SER contributes to affective empathic expression. Suppression and amplification produce consistent directional changes in expressed-empathy scores and support-strategy composition, with generally smaller changes in broader capabilities.

Critical strengths
  • A coherent chain from localization to independent validation: The DiD construction addresses input-side distress sensitivity and generic supportive language. Freezing the selected populations before held-out testing and evaluation on EPITOME and ESConv provides evidence beyond the discovery contrast itself.
  • Consistent bidirectional intervention evidence: Suppression and amplification change expressed empathy and support-strategy composition in consistent directions across four model families. Matched neuron controls and broader capability evaluations strengthen the case that SER contributes to response generation, rather than merely correlating with supportive text.
Critical weaknesses
  • Semantic validation of the generated conditions remains limited: The authors provide construction prompts, surface-form constraints, and rule-based checks, and acknowledge construct-validity limitations. However, Appendix C.4 reports exporting 100 response quadruplets for human inspection without reporting annotation outcomes. Independent checks of whether responses satisfy the intended conditions would strengthen the empathy-versus-support interpretation, especially given differences in how reassurance is grouped across experiments. This limits precise functional attribution, although independent-dataset results provide complementary evidence.
  • The understanding–expression dissociation is supported indirectly: Distress-related controls and broader emotion benchmarks provide useful selectivity evidence, and the authors appropriately claim only partial dissociation. Nevertheless, the experiments do not directly establish preserved understanding of the same users’ distress under SER suppression. The causal evidence is stronger for modulation of expressed empathy than for preserved distress understanding; this distinction should remain explicit.
Questions for authors
  1. Was the human inspection described in Appendix C.4 completed? If so, what proportion of responses satisfied the intended factors, and how was this assessed? How should readers reconcile the treatment of reassurance across the discovery, factorial, and ESConv analyses?
  2. Is there existing evidence that distress understanding remains intact under SER suppression on the same contexts where expression changes? If not, could the discussion distinguish the indirect evidence for partial dissociation from the directly demonstrated effect on expression?
3 comments
AuthorsOpenReview ↗

Response to Reviewer NvZP

We sincerely thank the reviewer for the careful reading and for raising two important concerns: the semantic validity of the factorial controls and the distinction between changing empathic expression and preserving distress understanding. We address both points below and have narrowed our claims accordingly.

  1. Human inspection and the treatment of reassurance

The human inspection described in Appendix C.4 was completed on the original seed-42 factorial sample, which contains 100 response quadruplets and 400 individual replies. Four independent annotations were collected for the same 400 replies. To keep the presentation concise, we report the averages across these four annotations:

Evaluation MetricAverage Rate
Reply-level conformity to the intended factors70.31%
Whole-quadruplet conformity14.0%
Both replies preserve the core facts97.5%
Neither reply adds unsupported facts97.5%
Both replies are fluent99.5%

The reply-level conformity rate measures whether an individual response contains the intended empathy/support factors. Whole-quadruplet conformity is stricter and requires all four responses in a quadruplet to satisfy their intended factors. The difference between these two measures is expected: a quadruplet fails the stricter criterion if any one of its four responses contains an unintended factor.

Although these four conditions are not completely semantically pure, they generally retain the original content and language quality. The averages above are computed from four independent annotations of the same 400 replies. They are not rates over 1,600 independent responses and do not represent an adjudicated consensus.

The different treatment of reassurance reflects the operational purpose of each experiment. The discovery contrast uses a broad supportive--neutral distinction that includes emotional acknowledgment, validation, reassurance, and interpersonal support. We retain the label “support” for the factorial conditions to remain consistent with the original protocol. Operationally, however, this factor represents additional supportive content, including reassurance and guidance, whereas the empathy factor denotes explicit empathic acknowledgment. ESConv retains its native strategy labels, under which Affirmation/Reassurance is grouped with affective support. These are distinct operational contrasts rather than a single universal taxonomy. We will clarify this distinction in the main text and Appendix C.4.

  1. Distress understanding when expression changes

We agree that a change in empathic expression does not by itself establish that distress understanding remains intact. The original full-prefill intervention acts during both prompt processing and response generation, so it cannot alone demonstrate unchanged prompt-side recognition.

To address this issue, we added a strict decode-only experiment on the same 1,200 source-disjoint EmpatheticDialogues inputs. The prompt is processed without SER scaling and its KV cache is retained; SER scaling is installed only immediately before response-token generation. Consequently, the binary distress-recognition and four-choice emotion-recognition readouts use exactly the unmodified prompt computation. Under both SER=0 and SER=2, both recognition readouts match baseline on every input, yielding 1,200/1,200 exact matches for each readout.

At the same time, expressed empathy changes in the expected bidirectional direction. The ER changes at SER=0 and SER=2 are as follows:

ModelER Change (SER=0)ER Change (SER=2)
Qwen-0.1595+0.1030
OLMo-0.0654+0.0489
Mistral-0.0868+0.0465
ChatGLM-0.1144+0.1070

All ER changes are based on 5,000 paired bootstrap replicates, and all corresponding 95% bootstrap intervals for ER exclude zero. This experiment directly demonstrates that SER can causally alter empathic expression while the prompt-side recognition computation is held fixed.

We note that decode-only isolates a generation-side expression effect but does not establish unchanged latent distress representations under full-prefill ablation. Full-prefill recognition equivalence is model- and dose-dependent: all models meet the criterion at SER=0.5; at SER=0, Qwen, OLMo, and ChatGLM do so, whereas Mistral does not. We therefore claim partial functional dissociation with interdependence, rather than complete representational independence.

We thank the reviewer again for the constructive feedback, which prompted these clarifications. In the revision, we will report the averaged human-audit results, distinguish the operational uses of reassurance across the three experiments, cite the corresponding appendix tables, and state the recognition--expression conclusion with the qualification above. We hope these clarifications address the reviewer’s concerns.

Thank you for the detailed response and the additional validation and experiments. These largely address my concerns, and I maintain my rating of 3: Weak acceptance.

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

This paper explores whether distress representation and empathic response realization can be mechanistically separated inside LLMs. Its contributions includes separating distress representation from empathic realization, a deconfounded identification paradigm, and validation across four model families. Experiments show encouraging generalization and consistent bidirectional causal effects from activation scaling while largely preserving general capabilities.

Critical strengths

The paper makes a genuinely useful distinction between input-side distress sensitivity, generic supportive style, and response-side empathic realization. This three-way confound is often glossed over in neuron-localization work, and framing it explicitly is a real contribution.

Critical weaknesses

1 Activation scaling is a coarse intervention that changes a dimension's contribution to every downstream readout and may induce out-of-distribution shifts. The paper does not report distribution-shift diagnostics (hidden-state norms, downstream activation distributions).

2 DiD validity requires that absent the treatment, distress and non-distress groups would differ in parallel. The paper reports no baseline distribution comparison beyond means and no leave-one-emotion-subcategory-out robustness analysis to show the effect is not driven by a single emotion class.

3 Four independent identifications with different set sizes and depth profiles support replicability, not identity. The singular "the mechanism" language is not warranted, and the paper should distinguish reproducibility from sameness.

Questions for authors

Q1. Does α-scaling match hard zeroing/mean-replacement in EPITOME effect, and can you report post-intervention hidden-state distribution-shift diagnostics to rule out off-distribution artifacts?

Q2. What evidence supports the DiD parallel-trends assumption, placebo tests with shuffled D/C labels, leave-one-emotion-out robustness, and beyond-mean comparisons across groups?

Q3. Does SER recruitment predict EPITOME beyond lexical baselines (emotion-word density, pronouns, length)? A null result would favor an "affective style" reading.

Q4. Would a control of neurons high on SERScore and highly responsive in non-distress contexts still underperform SER?

Q5. Can you report Wald–permutation agreement, fold-to-fold SERScore rank correlation, and sensitivity to the 1% threshold (0.5%, 2%)? Is 1% fixed a priori?

Q6. The provided Table 3 contains only headers. Could you supply baseline scores and paired deltas for all endpoints?

Q7. ER0 recruitment is negative in several models (e.g., −0.0377 for Qwen). What does sub-baseline recruitment mean mechanistically?

Q8. With a 61.2% acceptance rate, how do retained and discarded pairs differ in emotion category, length, and gold-response properties?

Q9. What is the EPITOME classifiers' agreement with human annotations, and do effects survive substitution with an alternative evaluator?

Q10. Do you claim a shared mechanism or replicability of an identification procedure across models? The "the mechanism" phrasing suggests the former.

7 comments
AuthorsOpenReview ↗

Response to Reviewer eSMh

We thank the reviewer for the detailed questions. We added intervention diagnostics, source-level robustness analyses, stronger controls for Supportive Empathic Response (SER) neurons, lexical prediction tests, and evaluator checks. We address all ten questions below and clarify the evidential scope of our conclusions.

Q1. Hard zeroing, mean replacement, and distribution shifts

Under our scaling operator for signed post-gating feed-forward network (FFN) activations at the input to the MLP down-projection, α=0 is exactly hard zeroing. We additionally evaluated independently calibrated mean replacement and 2× enhancement clipped to each target neuron’s baseline 0.5th–99.5th percentiles. Each model uses 768 EPITOME test inputs and 384 disjoint calibration inputs, with neuron sets and calibration statistics frozen before testing.

ModelMean-replacement ΔERClipped-2× ΔERClipped neutral-text PPL ratio
Qwen-0.5900+0.08041.010
Mistral-0.4825+0.06911.008
OLMo-0.2111+0.03940.999
ChatGLM3-0.2157+0.04610.986

All corresponding 95% paired bootstrap intervals for ΔER exclude zero. Mean replacement therefore reproduces the suppression direction, while bounded enhancement preserves the enhancement direction; we do not claim identical effect magnitudes. ER denotes Emotional Reaction, scored on the 0–2 scale. This endpoint differs from the EPITOME intervention score in Table 3, which averages the expected scores for ER, Interpretation (IP), and Exploration (EX).

We also measured residual/FFN norms, full-vector relative L2, cosine similarity, projected Mahalanobis atypicality, fixed-reference next-token KL/NLL, generation repetition/truncation, and neutral-text PPL. At α=0.5, ER still decreases in all models, with the following fixed-reference residual-state diagnostics:

ModelNorm change (%)Relative L2Atypical-layer change (pp)
Qwen-0.400.0752-0.06
Mistral-0.480.0893-0.07
OLMo+0.170.0932-0.06
ChatGLM3-0.660.0700-0.11

Atypicality uses a fixed 32-dimensional projection and an independently calibrated 99% threshold. These checks constrain specific artifact explanations rather than establish that every joint state remains in-distribution. In particular, Qwen clipped enhancement increases residual norm by 3.43% and projected atypical-layer fraction by 1.84 pp; OLMo mean replacement increases neutral PPL by approximately 30%. We report these exceptions alongside the favorable results.

Q2. DiD interpretation, label-shuffle placebos, emotion exclusions, and distributions

We agree that the assumptions behind the DiD interpretation should be explicit. Our DiD is a cross-sectional interaction: the supportive–neutral activation difference in distress contexts minus the corresponding difference in non-distress contexts. The relevant assumption is that nuisance effects introduced by rewriting are sufficiently comparable across the two context groups for the control contrast to remove generic-support effects. Matched contexts and response constraints help address this concern, but do not prove the assumption. We will clarify this interpretation rather than claim that a temporal parallel-trends assumption has been empirically verified.

We added three complementary checks, keeping discovery and held-out sources disjoint and averaging repeated turns within each source before inference.

First, for the original frozen SER populations, we shuffled distress/control labels at the held-out-source level 10,000 times. Second, we performed two leave-one-emotion-out (LOCO) analyses: retaining the frozen population and excluding each emotion from evaluation; and excluding each emotion from discovery, reselecting the population, then evaluating on independent held-out sources with that emotion also excluded.

ModelFrozen SER held-out DiD (95% CI)Two-sided label-shuffle pFrozen LOCO: positive CIs / exclusionsRediscovery LOCO: positive CIs / exclusions
Qwen0.01308 (0.01220, 0.01400)1/10,00130/3030/30
Mistral0.01014 (0.00954, 0.01074)1/10,00130/3030/30
OLMo0.02054 (0.01764, 0.02331)1/10,00130/3030/30
ChatGLM0.01290 (0.01202, 0.01381)1/10,00130/3030/30

The DiD and permutation results in this table use newly extracted activations for the original frozen SER sets; the rediscovery column concerns separately reselected sets. LOCO counts refer to strictly positive 95% bootstrap intervals. The permutation p-values reach the Monte Carlo resolution limit. These tests show that the observed interaction is unusual under source-label shuffling and remains positive after excluding any single emotion category; they do not independently validate the rewriting assumption.

Third, to address comparisons beyond group means, we summarize the source-level paired activation gap, δ = A(supportive) − A(neutral), using the original saved activation cache:

ModelContextSourcesMean δSDMedian δ25th–75th percentiles
QwenDistress24230.021760.020690.01698(0.00632, 0.03247)
QwenNon-distress22270.007450.011890.00402(0.00049, 0.01053)
MistralDistress14940.014040.009910.01249(0.00654, 0.02150)
MistralNon-distress13210.002730.007500.00131(-0.00192, 0.00547)
OLMoDistress1860.024120.017950.02409(0.00792, 0.03609)
OLMoNon-distress3180.003720.009980.00100(-0.00195, 0.00598)
ChatGLMDistress24230.023880.020800.01932(0.00887, 0.03400)
ChatGLMNon-distress22270.009800.011380.00720(0.00292, 0.01349)

These are descriptive distribution summaries, not confidence intervals. Median paired gaps are larger in distress contexts in all four models. We will also include the standard deviations and 5th/10th/25th/50th/75th/90th/95th percentiles for supportive activation, neutral activation, and their paired gap. Original-cache summaries and the newly extracted robustness results are reported separately; they are not pooled into a single numerical estimate. Together, these checks strengthen the robustness of the distress-conditioned activation interaction while retaining the stated construction assumption.

Q3. Prediction beyond lexical baselines

We evaluated 2,988 human-labeled EPITOME reference responses from 2,944 sources using grouped five-fold prediction. The lexical baseline includes log response length, emotion-word density, first-/second-person pronoun density, and question/exclamation features. The augmented ridge predictor adds frozen SER recruitment. Normalization and fitting use training folds only, with identical folds and regularization for both predictors. The table below reports prediction of human-rated ER.

ModelLexical RMSELexical + SER RMSEMSE improvement (95% source-bootstrap CI)
Qwen0.54620.51290.03531 (0.02628, 0.04385)
Mistral0.54620.51640.03168 (0.02194, 0.04072)
OLMo0.54620.50850.03980 (0.02964, 0.04962)
ChatGLM0.54620.50560.04275 (0.03335, 0.05279)

All four models also improve prediction of the human ER/IP/EX mean. Thus, SER provides incremental information beyond the specified lexical features. This analysis uses human dataset labels rather than intervention-scorer predictions and does not establish independence from every possible style feature.

Q4. High-SERScore, high-nondistress-response controls

Using discovery data only, we constructed two controls: neurons jointly high in SERScore and absolute non-distress response activation (joint-level), and neurons jointly high in SERScore and the non-distress supportive–neutral activation gap (joint-gap). Both quantities must be positive and in the within-layer top quartile. We exclude SER neurons and exactly match population size and layer counts; we do not claim exact matching of score distributions.

On 128 distress and 128 non-distress held-out contexts per model, SER exceeds both controls:

ModelSER−joint-level DiD (95% CI)SER−joint-gap DiD (95% CI)
Qwen0.0835 (0.0459, 0.1244)0.0839 (0.0665, 0.1016)
Mistral0.1178 (0.0750, 0.1606)0.1622 (0.1304, 0.1936)
OLMo0.1635 (0.1218, 0.2021)0.1882 (0.1535, 0.2227)
ChatGLM0.0597 (0.0213, 0.1005)0.0765 (0.0613, 0.0919)

These activation contrasts use frozen discovery-standard-deviation scaling. Paired free-generation tests on 768 EPITOME inputs per model also favor SER in all eight model–control ER comparisons under suppression and all eight under enhancement. This supports relative functional enrichment beyond these generic-support alternatives, rather than empathy exclusivity.

Q5. Wald–permutation agreement, fold stability, and selection size

We recomputed source-equal Welch/Wald and studentized permutation tests over all neurons, with 10,000 permutations and full-neuron BH correction for each test. We also calculated all ten pairwise SERScore rank correlations across the five original source folds.

ModelWald/permutation rejection-set JaccardOriginal SER passing permutation BHAll-neuron median Spearman rhoWithin-SER median rho
Qwen0.994199.7%0.8200.456
Mistral0.996099.3%0.8800.527
OLMo0.9830100.0%0.5720.094
ChatGLM0.993399.6%0.7720.553

Reselection at 0.5%, 1%, and 2% yields positive held-out 95% intervals in all 12 model–size combinations. Existing causal ablations at these sizes also reduce EPITOME in every model; these generations were not newly rerun. Population-level functional robustness is therefore stronger than exact within-SER rank stability, particularly in OLMo. The original 1% setting was fixed before downstream evaluation but was not preregistered; we do not claim it is uniquely optimal.

Reviewer_P9GfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary: The paper argues that recognizing user distress and expressing empathy rely on partially separable LLM mechanisms. It identifies sparse post-gating FFN units called Supportive Empathic Response (SER) neurons.

Contributions: A counterfactual difference-in-differences method isolates response-side empathy activity from input distress sensitivity and generic supportive language. SER neurons preferentially track affective empathy rather than instrumental support or distress intensity.

Validation: Across four model families, frozen SER sets generalize to held-out and independent datasets. Suppression reduces EPITOME empathy scores by 10.3%–31.9% and shifts responses toward instrumental support; amplification reverses these effects, with limited impact on reasoning and generation benchmarks. Factorial tests, matched-neuron controls, and sparse ablations further support selectivity and causality.

Critical strengths

The paper distinguishes between distress-related and Supportive Empathic Response neurons, offering a novel mechanistic perspective on empathetic dialogue in LLMs.

Critical weaknesses
  1. I have concerns about the paper’s motivation. The paper frames its central question as: "can distress representation and empathic response realization be mechanistically separated within LLMs?" However, based on the second paragraph of the introduction, I believe the more meaningful question is why recognized distress may fail to translate into an appropriate empathic response, and how this connection could be strengthened.

  2. The empirical evidence does not fully support the claimed separation. The paper primarily identifies and intervenes on empathy-related SER neurons, but provides no comparable identification or causal analysis of distress-representation neurons. The current results show that manipulating SER neurons changes expressed empathy, but do not establish how SER neurons relate to distress representations or whether the two mechanisms are genuinely separable.

  3. I am concerned about the supportive–neutral construction. Because these responses are controlled rewrites subsequently processed through teacher forcing, it is unclear whether the identified activations reflect mechanisms used during natural response generation or merely neural responses to the constructed text. This distinction is important for interpreting SER neurons as a generative mechanism.

  4. The presentation and writing require further polishing. For example, the spacing between the captions of Figures 1 and 2 and the following text is too narrow, while the notes below Tables 1–3 appear visually abrupt and insufficiently integrated into the layout. These formatting issues reduce readability and make the presentation feel less polished.

Questions for authors

See Weaknesses.

13 comments
AuthorsOpenReview ↗

Response to Reviewer P9Gf

We thank the reviewer for the careful reading and helpful points on motivation, dissociation, teacher forcing, and presentation. We address each point below and clarify the evidential scope of our claims.

1. Motivation: strengthening the recognition--expression link

We agree that the motivation should make the link between distress representation and empathic response realization explicit. Our work addresses this issue by testing whether a specific response-side component can strengthen empathic expression after distress recognition is already correct and input-side distress sensitivity is controlled.

We directly tested this connection using distress inputs that were correctly recognized at baseline from 1,200 source-disjoint EmpatheticDialogues messages. SER amplification increased expressed-empathy scores across all four models. Adding SER amplification after a correct-state distress reminder produced additional gains over the reminder alone:

ModelnΔER:BoostΔER:Boost + Distress-reminder
Qwen528+0.5285 (0.4812, 0.5774)+0.2574 (0.2249, 0.2905)
Mistral486+0.1235 (0.0977, 0.1503)+0.0148 (0.0056, 0.0259)
OLMo513+0.1523 (0.1231, 0.1827)+0.0778 (0.0567, 0.1001)
ChatGLM564+0.2275 (0.1915, 0.2659)+0.2219 (0.1908, 0.2541)

Both contrasts use 2× SER scaling. ER denotes Emotional Reaction on the 0–2 scale. Parentheses give 95% paired bootstrap intervals.

Parentheses report 95% paired bootstrap intervals, all of which exclude zero. Post-enhancement binary recognition accuracy remained at 99.2--100%, with all 90% accuracy-change intervals within the prespecified ±3 percentage-point equivalence margin.

These results show that SER modulation can further strengthen empathic expression after distress recognition is already correct. Thus, our findings connect the recognition and expression stages by identifying a causally adjustable response-side component, while retaining the more cautious conclusion that the two processes are partially dissociable but functionally interdependent.

2. Scope of the dissociation claim

We agree that we do not identify a complete distress-representation circuit with a symmetric causal analysis. Accordingly, we do not claim two fully separable mechanisms. We identify model-specific response-side SER populations that causally influence expressed empathy, while recognition readouts remain equivalent under some intervention strengths.

As a control for input-side distress sensitivity, we construct matched distress-control populations from prompt-level distress--control activation differences, excluding SER neurons and matching neuron count and layer distribution. Across all four models, SER shows stronger held-out selectivity than the distress-control population:

ModelSER DiDDistress-control DiDSER Pair AccuracyDistress-control Pair Accuracy
Qwen0.01370.00340.86670.5277
OLMo0.02040.00150.93500.4100
Mistral0.01110.00060.93190.3692
ChatGLM0.01360.00100.90200.5007

Across independent EPITOME and ESConv evaluations, SER recruitment was more closely associated with affective empathic expression than with distress magnitude.

The full-prefill behavioral recognition analysis is model- and dose-dependent. At SER=0.5, all four models satisfy the behavioral recognition-equivalence criterion while ER decreases in all four. Under full ablation, behavioral recognition remains equivalent for Qwen, OLMo, and ChatGLM but not Mistral. We therefore interpret the evidence as partial functional dissociation with interdependence, rather than complete preservation of all distress representations or complete mechanistic independence.

3. Teacher forcing versus natural generation

The supportive--neutral rewrites are used for discovery and localization, not as the sole evidence for generation. SER identities are frozen before all downstream analyses. Causal interventions are then applied during free generation on 3,023 independent EPITOME contexts and 2,518 ESConv contexts. Suppression lowers expressed empathy and amplification raises it in all four model families; ESConv affective--instrumental strategy margins change in the corresponding direction.

We also tested whether SER recruitment adds information beyond surface lexical features. In source-grouped five-fold prediction on 2,988 independently human-labeled EPITOME responses from 2,944 sources, adding frozen SER recruitment improves ER prediction in every model:

ModelLexical RMSELexical + SER RMSEMSE Improvement (95% CI)
Qwen0.54620.51290.03531 (0.02628, 0.04385)
Mistral0.54620.51640.03168 (0.02194, 0.04072)
OLMo0.54620.50850.03980 (0.02964, 0.04962)
ChatGLM0.54620.50560.04275 (0.03335, 0.05279)

These results do not prove that every rewrite is semantically perfect, but show that the downstream signal goes beyond the tested lexical features and transfers to model-generated responses.

4. Presentation

We agree that the presentation can be improved. We will increase caption spacing, integrate the notes below Tables 1--3 into the captions, and improve table hierarchy without changing the results.

We thank the reviewer for the constructive feedback. We will clarify the motivation, distinguish discovery localization from free-generation validation, and state the conclusion as causal evidence for response-side empathic realization with qualified evidence for partial functional dissociation.

I thank the authors for their detailed responses. However, after reviewing both the response and the revised manuscript, I remain unclear about the alignment between the paper’s motivation (L53) and its empirical evidence.

In my view, the current evidence is asymmetric. The paper identifies and causally manipulates a response-side SER population. However, it does not independently identify or causally validate a distress-recognition mechanism. This creates a mismatch with the title, “Understanding Distress Is Not Expressing Empathy.” The evidence more directly supports the narrower conclusion that manipulating a response-side empathy-related population can alter expressed empathy.

Therefore, I maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.