HEART: A Unified Benchmark for Assessing Humans and LLMs in Emotional Support Dialogue

TL;DR

HEART presents a first-of-its-kind benchmark that directly pits humans and large language models in multi-turn emotional-support dialogue, evaluated with blinded human raters and LLM evaluators across five interpersonal axes. By leveraging a large, adversarially enriched dataset based on ESConv and a rigorous pairwise evaluation framework, the study shows frontier LLMs approaching or exceeding average human performance on perceived empathy while humans retain advantages in adaptive nuance and boundary setting. The findings reveal a meaningful convergence in evaluative criteria between humans and models, yet persistent gaps in dynamic, context-sensitive support—especially under resistance—highlight the need for domain-specific tuning and safety-conscious deployment in real-time, affective AI applications. Overall, HEART provides an empirical scaffold for understanding how affective conversational competence scales with model capability and how to combine human judgment with AI assistance to optimize emotionally intelligent interactions.

Abstract

Supportive conversation depends on skills that go beyond language fluency, including reading emotions, adjusting tone, and navigating moments of resistance, frustration, or distress. Despite rapid progress in language models, we still lack a clear way to understand how their abilities in these interpersonal domains compare to those of humans. We introduce HEART, the first-ever framework that directly compares humans and LLMs on the same multi-turn emotional-support conversations. For each dialogue history, we pair human and model responses and evaluate them through blinded human raters and an ensemble of LLM-as-judge evaluators. All assessments follow a rubric grounded in interpersonal communication science across five dimensions: Human Alignment, Empathic Responsiveness, Attunement, Resonance, and Task-Following. HEART uncovers striking behavioral patterns. Several frontier models approach or surpass the average human responses in perceived empathy and consistency. At the same time, humans maintain advantages in adaptive reframing, tension-naming, and nuanced tone shifts, particularly in adversarial turns. Human and LLM-as-judge preferences align on about 80 percent of pairwise comparisons, matching inter-human agreement, and their written rationales emphasize similar HEART dimensions. This pattern suggests an emerging convergence in the criteria used to assess supportive quality. By placing humans and models on equal footing, HEART reframes supportive dialogue as a distinct capability axis, separable from general reasoning or linguistic fluency. It provides a unified empirical foundation for understanding where model-generated support aligns with human social judgment, where it diverges, and how affective conversational competence scales with model size.

Figures (7)

• Figure 1: HEART dimensions and rubric. Venn-style overview of the five evaluation axes—Human alignment, Empathic Responsiveness, Attunement, Resonance, and Task-following—with representative sub-criteria used by raters.
• Figure 2: Example pairwise comparison evaluation. Pairwise comparison task used for human judges, where two responses (A and B) are rated using a five-point preference scale. This setup allows direct comparison between human and LLM completions across identical dialogue histories.
• Figure 3: Example evaluation. Each dialogue presents a multi-turn emotional support exchange between a seeker and supporter (left). Two candidate responses are evaluated by both human and LLM judges. LLM evaluators provide graded comparative scores on the five HEART dimensions (“1+” to “1++++”) together with chain-of-thought rationales, which are aggregated into Elo-style scores and coded for themes in parallel with human rationales.
• Figure 4: HEART leaderboard with per-dimension percentiles. Each row shows a model’s percentile on Human alignment (H), Empathic Responsiveness (E), Attunement (A), Resonance (R), and Task-following (T), alongside overall Elo. Darker colors indicate higher percentile performance. Humans show relatively balanced performance across dimensions, whereas models display characteristic profiles associated with model family and alignment. Overall win rate reflects the raw proportion of pairwise A/B comparisons in which a system’s response was preferred on the primary HEART judgment.
• Figure 5: Performance rises with latency used as a proxy for capacity. Scatter of HEART Elo versus median time-to-first-token (log scale), derived from AA Analysis artificialanalysis2024. Each point represents a model variant, colored by provider. The dashed horizontal line marks the field mean. See Section \ref{['sec:latency-requirements']} for discussion of real-time latency constraints. Latency is shown on a log scale to accommodate variation across models.