Detection != Reliable Control: Decodable Empathy Directions Yield at Most Partial Shifts in Automated Empathy Scores

Motivation

• Decodability is insufficient for causal claims because probe accuracy need not indicate genuine encoding, model use, or reliable steering.Prior work also reports brittle steering vectors and interpretability illusions.Source: §1 · p. 1
• Empathy is tested as a high-stakes, multidimensional construct with cognitive Recognition and affective Resonance facets.Recognition denotes communicating inferred understanding of a seeker’s experience, while Resonance concerns affective feeling-related responses.Source: §1 · p. 1

Findings

• Recognition and Resonance remain decodable after residualizing against a sentence-embedding-derived surface score.This establishes detection beyond the tested surface-text control, not reliable intervention control.Source: §1 · p. 2
• Affective steering is a graded, sub-maximal automated-metric shift: Qwen Resonance at α=0→+8 rises +0.29, approximately 26% of the natural gap.The shift is facet-differentiated in Qwen and Llama, but it is measured by classifiers rather than established in human perception.Source: §1 · p. 2
• Additive Recognition steering is unmeasurable because a within-domain control shows insufficient cognitive-instrument sensitivity, whereas Gemma Recognition ablation lowers the classifier score.The resulting cognitive additive null is not established as a clean absence of effect.Source: §1 · p. 2

Probes

• Probe directions are differences between high- and low-level EPITOME class-mean residual-stream activations, evaluated layerwise with Cohen’s d.For each facet, the procedure samples 100 Level 2 and 100 Level 0 items in Client/Counselor exchanges and extracts last-token, pre-generation activations.Source: §2 · p. 2

Surface-text residualization

• Surface-text residualization regresses probe projections on a cross-validated sentence-embedding surface score and recomputes residualized effect sizes.Each facet’s peak layer is selected by the maximum residualized d, making residualization a detection control.Source: §2 · p. 2

Causal steering

• At each facet’s peak layer, additive steering scales the probe direction by its projection standard deviation across α ∈ [−8, 8] and generates deterministic responses.The intervention uses 50 prompts per condition, including emotional and neutral prompts, while a random-direction floor addresses differences in raw perturbation magnitude.Source: §2 · pp. 2–3

Evaluation

• A blind GPT-4o judge rates cognitive and affective empathy separately, while concept-specific effects must exceed the other dimension and a paired random-direction floor.The analysis uses bootstrap 95% confidence intervals, Bonferroni-corrected lexical analyses, and degeneration checks.Source: §2 · p. 3

Production-side direction

• The study also estimates a response-token Recognition direction, compares its cosine with the reading-time probe direction, and steers at its own peak layer.This separates directions derived from model reading-time activations from those derived from generated response-token activations.Source: §2 · p. 3

Directional ablation

• Directional ablation removes the peak-layer unit direction from every layer and token position during generation, with matched random-direction effects tested using paired sign-flip permutation and Wilcoxon tests.The residual stream is treated as one shared space across layers, and the random-direction floor aggregates five directions per prompt.Source: §2 · p. 3

Output-change diagnostic

• Output-change diagnostics compare α=0 and α=+8 generations using MPNet cosine, random- and cross-prompt floors, and a unique-token-ratio degeneration threshold.Lower cosine indicates larger text change, while collapse is flagged below a unique-token ratio of 0.5.Source: §2 · p. 3

Instruments and controls

• Every intervention is evaluated with two additional automated instruments and an emotional-vs-neutral positive control, expanded to 40 neutral prompts for cognitive sensitivity.The cognitive control’s original 10 neutral prompts implied a minimum detectable d of approximately 1.0, motivating 30 added Dolly factual prompts.Source: §2 · pp. 3–4

Detection

• Both Recognition and Resonance remain decodable after surface-text residualization, with model-dependent geometric relations between directions.Residualized effect sizes remain substantial in every model; Recognition peaks later than Resonance, and their geometry is descriptive rather than causal.Source: §3 · p. 4

Cognitive steering

• Recognition steering substantially changes text while leaving cognitive scores flat, but within-domain controls show the cognitive instrument cannot resolve the relevant differences.The additive cognitive result is therefore unmeasurable rather than a clean null; Qwen lacks cognitive classifier range, while Llama and Gemma show only coarse range.Source: §3 · pp. 4–5, Figure 1 · p. 6

Affective steering

• +0.29: Qwen’s classifier-measured affective score rises from α=0 to +8, approximately 26% of its natural emotional-vs-neutral gap.The corresponding gains are +0.11 for Llama and +0.05 for Gemma; the full-span gains are +0.43/+0.25/+0.06.Source: §3 · p. 4
• Resonance produces an ordered, facet-specific affective shift in Qwen and Llama, but not Gemma, with no Recognition contrast surviving.The direct between-direction contrast is the primary specificity test; Gemma’s contrast is −0.01 with p=0.83.Source: §3 · p. 4

Interpretation

• Automated affective shifts cannot be identified as human-perceived empathy changes because the EPITOME-ER classifier is built from EPITOME-ER labels.The paper states that separating added emotional register from added empathy requires human evaluation.Source: §3 · p. 5

Ablation

• Gemma Recognition ablation lowers the classifier cognitive score from 0.31 to 0.11, and the drop survives adjustment for response length.The adjusted classifier effect is −0.19 with a 95% CI of [−0.35, −0.03], while affective scores remain at ceiling.Source: §3 · pp. 6–7

• Prior work links affective concepts to decodable directions and activation steering, but probe accuracy alone does not establish genuine encoding or model use.Source: §4 · p. 7
• The paper distinguishes EPITOME communicative empathy under global interventions from related work on behavioral or localized empathy steering.Cadile studies an “empathy-in-action” direction, while Chebrolu et al. and Tak et al. report localized interventions.Source: §4 · p. 7
• Affective classifier scores cleanly separate emotional from neutral prompts in all three models, with Cohen’s d=3.1–4.8.The affective positive control uses 40 emotional and 40 neutral responses.Source: Figure 3 · p. 8
• Cognitive classifier distributions overlap heavily: Llama and Gemma are moderately separable, whereas Qwen is exactly coincident.Cognitive d values are 0.62 for Llama, 0.67 for Gemma, and 0.00 for Qwen.Source: Figure 3 · p. 8
• The two LLM judges likewise agree more reliably on affective than cognitive scores, supporting measurement checks before interpreting cognitive null results.Inter-judge Krippendorff α is 0.70–0.84 for affective scores versus 0.24/0.44/−0.02 for cognitive scores across Llama, Gemma, and Qwen.Source: Figure 3 · p. 8, §4 · p. 8

• Recognition and Resonance directions remain decodable after residualizing against a surface score, but decodability does not yield reliable global control.Source: §5 · p. 8
• Resonance steering produces a model-dependent, sub-maximal shift in an automated affective metric, whereas additive Recognition steering is unmeasurable.Source: §5 · p. 8
• Only Gemma shows a length-adjusted ablation drop in the classifier’s cognitive score.Source: §5 · p. 8
• The paper does not establish a matching human-perceived empathy change, and a six-rater panel could not adjudicate the Qwen α=0 versus +8 contrast.Source: §5 · p. 8

• Each facet uses modest EPITOME item counts, and small cross-model ablation samples leave smaller necessity effects unresolved.Source: Limitations · p. 9
• Cognitive range is inconsistent across instruments, and the classifier cannot resolve the within-domain differences probed by additive steering.Accordingly, additive cognitive steering is interpreted as unmeasurable rather than null.Source: Limitations · p. 9
• Human and within-domain checks are bounded by hand-constructed items and one constructor’s construal of “understanding.”Source: Limitations · p. 9
• The affective effect is established only on automated instruments that do not fully agree, so it is framed as control of a steerable automated metric rather than human-perceived empathy.Source: Limitations · p. 9
• EPITOME labels the counselor turn while analysis uses Client+Counselor context, leaving a residualized but unresolved context confound.Source: Limitations · p. 9
• All interventions and measurements concern one generated response to one prompt, leaving multi-turn and long-context steerability unestablished.Source: Limitations · p. 9

• Six adult volunteers participated in a 10–15-minute pairwise-comparison text-rating task after reviewing consent information.Source: Ethics Statement · p. 9
• The study uses emotional-support prompts from ESConv, factual prompts from databricks-dolly-15k, and EPITOME empathy annotations.Source: Ethics Statement · p. 9
• The paper cautions against deploying activation-level empathy steering in emotional-support settings without evidence of reliable human-perceived change.The authors specifically warn that automated metrics may move without demonstrated human-perceived change.Source: Ethics Statement · p. 10, A Example Steered Generations · p. 11
• The retrained EPITOME classifier is a Sharma et al. bi-encoder with an empathy-identification head, trained on a stratified 80/20 held-out split.The test set contains n=617 items per head, using the same scoring configuration.Source: B Discriminative classifier: held-out… · p. 12
• The cognitive IP head predicts level 1 for 0/617 test items, making it effectively binary and limiting resolution for within-domain cognitive contrasts.Its macro-F1 is 0.558, while accuracy is 0.822 on the dominant 0/2 levels.Source: B Discriminative classifier: held-out… · p. 12, Table 3 · p. 12
• The positive control shows strong affective separation across instruments, but cognitive separation is inconsistent across judges and the classifier.Affective Cohen’s d is d≥2.6; cognitive results vary by instrument and model, with marginal classifier results for Llama and Gemma.Source: C Per-instrument positive control · p. 12, Table 4 · p. 12

• Cross-model ablation uses n=15, which can detect only effects ≳1 point at the observed standard deviations.Gemma Recognition is verified at n=50, with minimum detectable dz≈0.40, corresponding to ≈0.16 points at the within-prompt difference SD ≈0.39.Source: E Sample sizes and pilot attenuation · p. 13

• Six volunteer raters first had to reliably order hand-built calibration pairs before steered-output nulls could be interpreted.Source: F Human panel: a sensitivity-gated check · p. 13
• 21/47 committed understanding trials selected the higher-understanding response, also near chance with binomial p≈0.56.No rater cleared the understanding gate individually.Source: F Human panel: a sensitivity-gated check · p. 13
• The steered pairs showed no consistent direction, with 16 baseline, 19 “same,” and 13 steered responses.The sign test was p=0.71, but the failed sensitivity gate makes this result uninterpretable as a null.Source: F Human panel: a sensitivity-gated check · p. 13

Cross-facet steering

• The cross-facet matrix tests both steering directions against both classifier heads using paired α=0→+8 changes.This design distinguishes a matched-facet effect from a generic increase across EPITOME-related scores.Source: G Specificity controls: cross-facet… · p. 13
• Adding Recognition never raises ER, whereas adding Resonance raises ER and lowers IP in Qwen and Llama, indicating facet-differentiated rather than generic inflation.Gemma shows no reliable IP change, and Qwen Recognition→IP is uninterpretable because Qwen has no cognitive range.Source: G Specificity controls: cross-facet… · pp. 13–14, Table 5 · p. 14

Raw-norm floor

• At matched raw L2 norm, Recognition text change is comparable to random-direction means, with MPNet cosine 0.86/0.78/0.78 versus 0.84/0.82/0.79.The affective score remains direction-specific at matched norm: only Resonance clears the random floor.Source: G Specificity controls: cross-facet… · p. 14

H Response-token direction and all-layer ablation

• Response-token directions are near-orthogonal to read-time directions and show no large steering or ablation effects across the three models.The sole departure is Gemma Recognition ablation.Source: H Response-token direction and… · p. 14, Table 6 · p. 14
• Gemma Recognition is the sole reported departure from the generally weak response-token and all-layer intervention effects.Table 6 summarizes this exception alongside the broader dissociation.Source: H Response-token direction and… · p. 14, Table 6 · p. 14

I Length control for the Gemma ablation

• −0.40 to −0.13: the GPT-4o judge’s cognitive drop becomes statistically nonsignificant after controlling for length and prompt.The judge rewards length by +0.0067 points/word, while the affective drop changes from −0.24 to −0.003.Source: I Length control for the Gemma ablation · p. 15, Table 7 · p. 15
• −0.199 to −0.193: the classifier’s cognitive drop remains after length adjustment, with a clustered-SE 95% CI of [−0.35, −0.03].The classifier’s length slope is approximately −0.003 E[level]/word and nonsignificant under prompt-clustered standard errors.Source: I Length control for the Gemma ablation · p. 15, Table 7 · p. 15
• The two instruments dissociate sharply after adjusting for completion length and prompt fixed effects.Length is treated as downstream of the intervention and therefore as a potential mediator.Source: I Length control for the Gemma ablation · p. 15, Table 7 · p. 15