Multi-turn Empathy Conversations

• Multi-turn empathy conversations are defined as dialogues where agents repeatedly interpret and express empathy over multiple turns, emphasizing emotional resonance and contextual understanding.
• Benchmark datasets like SENSE-7 and PosEmoDial provide multi-dimensional annotations to evaluate empathy through metrics such as emotion tracking, validation, and context remembrance.
• Advanced models employ hierarchical encoders, adaptive pooling, and reinforcement strategies to dynamically track and integrate affective cues across extended conversational turns.

Multi-turn Empathy Conversations are a class of human–human or human–AI dialogues that require agents to repeatedly interpret, express, and adapt empathic behaviors across a sequence of conversational turns. Unlike single-turn exchanges, these interactions demand temporal coherence, context tracking, affective adaptation, and discourse strategy variation over multiple messages. Multi-turn empathy is fundamental to domains such as psychological counseling, social support, coaching, and collaborative work, and informs the development and assessment of both experimental and real-world conversational AI models.

1. Core Definitions, Taxonomies, and Data Resources

Empathy in conversation is operationalized at multiple levels: (a) affective resonance (emotional mirroring), (b) cognitive understanding (perspective-taking, goal inference), (c) response appropriateness (validation, advice, questioning), (d) prosocial expression (willingness to help), (e) interest (active inquiry), (f) contextual understanding (referencing user history/culture), and (g) relational continuity (recalling previous exchanges for rapport maintenance). SENSE-7 explicitly codifies these as seven dimensions, rated by users per turn on a discrete scale (Suh et al., 19 Sep 2025). Annotation schemas further distinguish between speaker self-reported empathy, partner other-reports, and third-party expert turn-by-turn assessments of empathy quality, self-disclosure, and emotion (Omitaomu et al., 2022).

Several benchmark datasets enable empirical study:

Dataset	Dialogs	Annotations	Domain
Empathic Conversations (Omitaomu et al., 2022)	≈1,200	Self/other/3rd-party empathy, emotion, disclosure	News-primed, dyadic
SENSE-7 (Suh et al., 19 Sep 2025)	672	7-dim user empathy, traits, context	Human-AI, open-domain
PosEmoDial (Wang et al., 2022)	≈820k	Automatic emotion trajectory, positive elicitation	Social web forums
SoulChatCorpus (Chen et al., 2023)	2.3M	Implicit empathy strategies via ChatGPT-guided dialog	Psychological support

These resources are complemented by specialized sets for mental health (MHSD (Chandra et al., 26 May 2025)), emotional support (ESConv (Bai et al., 21 May 2025)), and conversation-level empathy/affect prediction tasks (WASSA tracks (Pereira et al., 2024, Singh et al., 2024)).

2. Model Architectures and Contextualization Strategies

The architecture of multi-turn empathy models reflects the complexity of dialog context, emotion tracking, and multi-dimensional empathy estimation.

• Contextual encoding: Leading models (EmpTransfo (Zandie et al., 2020), ConText (Pereira et al., 2024)) process utterance k-history windows with segment or role delimiters and feed the expanded context to deep pretrained Transformers, extracting pooled representations for each utterance or the full dialog. Hierarchical attention encoders (HRAN) in (Xie et al., 2019) and graph convolutional models in (Altarawneh et al., 2023) provide fine-grained handling of word/utterance-level dependencies and inter-speaker relationships, especially for long or multi-party interactions.
• Emotion and empathy conditioning: Emotional state is commonly integrated as a per-utterance embedding, using either gold or predicted emotion classes (Zandie et al., 2020, Xie et al., 2019), VAD lexicon mapping (Wang et al., 2022), or predicted continuous emotion trajectories (Altarawneh et al., 2023). Some systems track “feature transitions” across turns—explicitly modeling changes in emotion, salient keywords, or semantic meaning (Emp-RFT (Kim et al., 2022)). Adaptive pooling or attention is often recency-weighted to avoid excessive influence from stale context (Altarawneh et al., 2023).
• Empathy prediction and generation heads: Most systems employ regression or classification heads for empathy and emotional traces on pooled representations (Pereira et al., 2024, Singh et al., 2024). Generation models utilize conditional Transformers (PLATO-2, BART, GPT-style) with losses targeting not only standard negative log-likelihood but also auxiliary objectives that enforce emotion guidance, empathy class, positive valence shift, or preference-based ranking (EmPO (Sotolar et al., 2024), PEGE (Wang et al., 2022), DPO (Sotolar et al., 2024)).

3. Specialized Training Objectives and Reinforcement Frameworks

Multi-task and reinforcement objectives are essential to align deep models with the complex, temporally extended nature of empathy.

• Multi-task objectives: Composite losses combine empathy classification, emotion regression, and self-disclosure (e.g., $$L = L_\mathrm{empathy} + \lambda L_\mathrm{valence} + \mu L_\mathrm{disclosure}$$ in (Omitaomu et al., 2022)). Models such as EmpTransfo employ output heads for next-token, next-utterance, and next-emotion, supporting both response and affect prediction (Zandie et al., 2020).
• Emotion/affect trajectory tracking: For positive emotion elicitation, the loss incorporates explicit valence-distance terms with turn-indexed weighting ($L_{peg}$ in (Wang et al., 2022)), regularizing early-turn empathy and late-turn elicitation in temporal alignment with dialog structure.
• Preference and RL-based alignment: Modern work applies preference optimization, e.g., Direct Preference Optimization (DPO) in EmPO, where models are trained to rank human or theory-grounded empathetic responses above non-empathic or affect-incongruent alternatives (Sotolar et al., 2024). Reinforcement learning–from–quality reward (e.g., PsychoCounsel) plus cross-turn discourse move diversity (KL-divergence on tactic distributions, as in MINT (Zhan et al., 13 Apr 2026)) provides empirical improvements in both empathy scores and tactic variety across turns.
• Adapters for context-specific empathy: Parameter-efficient fine-tuning (LoRA-defined adapters) specializing in context/task clusters notably improve the capacity of models to sustain target empathy levels across turns and tasks, outperforming prompt-only approaches (Empathetic Expert Adapters (Shayegani et al., 5 Nov 2025)).

4. Multi-Turn Evaluation Protocols and Metrics

Evaluation protocols reflect the multidimensional character of multi-turn empathy.

• Automatic metrics: Standard text overlap (BLEU, ROUGE), semantic similarity (BERTScore), and perplexity are commonly reported (Chen et al., 2023, Sotolar et al., 2024), but show limited correlation with empathy perception (Sotolar et al., 2024).
• Empathy/affect scoring: Datasets with turn-wise empathy and emotional traces enable metrics such as accuracy, macro-F1, and Pearson correlation against ground-truth or rated targets (Omitaomu et al., 2022, Pereira et al., 2024, Singh et al., 2024). Feature-specific metrics such as PEG-Score (final user positive valence gain), E-Score (early-turn empathy), and combined PEGE-Score (Wang et al., 2022) directly target intended affective shift.
• Discourse-level and multi-turn metrics: Recent research tracks tactic “stickiness” (repeated use of a supportive move across turns) (Zhan et al., 13 Apr 2026), cross-turn tactic novelty (KL-divergence), relational continuity, within-k accuracy, and dialogue-level mean empathy (Suh et al., 19 Sep 2025, Shayegani et al., 5 Nov 2025).
• Human (and user) evaluation: Direct human and end-user ratings remain authoritative for dimensions such as naturalness, emotional appropriateness, turn-wise empathy, overall satisfaction, and engagement (Suh et al., 19 Sep 2025, Omitaomu et al., 2022, Chen et al., 2023, Jia et al., 26 Jan 2026). Inter-annotator agreement, Cronbach's α, Cohen's κ, and other reliability statistics are reported.

5. Behavioral Analysis and Observed Challenges in Multi-Turn Empathy

Empirical analysis reveals both the promise and current limitations of multi-turn empathy models.

• Temporal context dependency: All models, regardless of backbone, suffer performance loss as the number of dialog turns grows (typically >5). Macro-F1 for empathy drops ≈1.5% per additional turn in (Omitaomu et al., 2022); diagnostic and affect scores decline similarly in longer sequences (Chandra et al., 26 May 2025).
• Discourse rigidity and tactic repetition: LLMs exhibit near-doubling of discourse move stickiness compared to humans (0.50–0.56 vs. 0.27), indicating overuse of the same supportive strategies across turns, which undermines perceived support variety and re-engagement (Zhan et al., 13 Apr 2026).
• Contextual/individual alignment: Trait-level user differences (Big Five, IRI) explain systematic variance in empathy production and perception; models aligned to user persona (agreeableness, literacy) perform better (Omitaomu et al., 2022, Chandra et al., 26 May 2025, Suh et al., 19 Sep 2025). Explicit user modeling, trait conditioning, and real-time empathy calibration are recommended.
• Robustness/continuity breakdown: A single “poor turn” (e.g., generic reply, missed context) produces a statistically significant drop in subsequent empathy and overall user engagement, emphasizing the cost of local failures in long contexts (Suh et al., 19 Sep 2025).
• Qualitative tradeoffs: Structurally enforced emotional arcs (VADER slopes, target trajectory correlation) are necessary but insufficient; real empathetic listening requires open-ended feeling exploration, probing, and genuine curiosity—qualities often lacking in large-scale LLM output (Knob et al., 3 Jul 2025). Multiple strategy use per turn (CUS) is critical for actionable and user-centered support (Bai et al., 21 May 2025).

6. Best Practices, System Design Recommendations, and Future Directions

Developing robust, user-centered multi-turn empathy systems demands integration of methodological advances and domain insights.

• Context-aware encoding: Encode recent conversational history (4–8 prior turns) with advanced segmenting and role marking. Maximize history coverage without exceeding transformer or context window limitations (Pereira et al., 2024, Chen et al., 2023).
• Emotion-adaptive and feature-transition modeling: Track and inject emotion, disclosure, and key features at each turn. Use transition-aware networks (Emp-RFT (Kim et al., 2022)) and self-dependency/recency-weighted mechanisms to maintain affective state and semantic shifts (Altarawneh et al., 2023).
• Discourse diversity optimization: Employ RL objectives that explicitly reward cross-turn behavioral variety, anchored by quality models (Zhan et al., 13 Apr 2026). Implement strategy sequencing or content planning to generate multi-strategy responses per turn (Bai et al., 21 May 2025).
• Context and user/persona alignment: Integrate user profile data and psychometrics into model inputs; adapt empathy patterns to the task context via expert adapters and preference modeling (Shayegani et al., 5 Nov 2025, Markó et al., 2024).
• Dynamic calibration and repair routines: Monitor per-turn empathy metrics, turn-level user ratings, and conversational engagement; trigger empathic repair sequences when breakdown is detected (Suh et al., 19 Sep 2025).
• Evaluation and continuous improvement: Combine scalable preference-optimized metrics, LLM-as-judge ratings, and direct human user feedback; adjust model optimization and data patterns in response to observed performance gaps (Sotolar et al., 2024, Shayegani et al., 5 Nov 2025).

Key open directions include multimodal integration (vocal, facial cues), reinforcement learning from richer user feedback, global discourse-level planning, longer-term memory for relational continuity, escalation/safety for critical tasks, and fine-grained, culture-aware empathy adaptation across diverse users and applications.