Large language models (LLMs) produce responses rated as highly empathic in single-turn settings (Ayers et al., 2023; Lee et al., 2024), yet they are also known to be formulaic generators that reuse the same lexical patterns, syntactic templates, and discourse structures across tasks (Jiang et al., 2025; Shaib et al., 2024; Namuduri et al., 2025). Less attention has been paid to whether this formulaicity extends to the level of discourse moves, i.e., what a response does for the person it is addressing. This question is especially consequential for empathic dialogue, where effective support demands not just a kind response at one moment but varied strategies as a conversation unfolds (Stiles et al., 1998). Indeed, prior work shows that LLMs reuse the same tactic sequences more than human supporters in single-turn settings (Gueorguieva et al., 2026). We extend this analysis to multi-turn conversations and find that the rigidity compounds: once a tactic appears in a supporter turn, LLMs reuse it in the next at nearly double the rate of humans (0.50-0.56 vs. 0.27). This pattern holds across LLMs serving as supporters in real emotional support conversations, and is invisible to standard similarity metrics. To address this gap, we introduce MINT (Multi-turn Inter-tactic Novelty Training), the first reinforcement learning framework to optimize discourse move diversity across multi-turn empathic dialogue. The best MINT variant combines an empathy quality reward with a cross-turn tactic novelty signal, improving aggregate empathy by 25.3% over vanilla across 1.7B and 4B models while reducing cross-turn discourse move repetition by 26.3% on the 4B model, surpassing all baselines including quality-only and token-level diversity methods on both measures. These results suggest that what current models lack is not empathy itself, but the ability to vary their discourse moves across a conversation.
Figures & tables
Figure 1: As the seeker’s needs evolve, vanilla LLMs recycle a narrow tactic set, whereas Mint adapts its discourse moves across turns.
Figure 2: Left: Tactic prevalence (percentage of turns containing each tactic). LLMs heavily overuse advice ( 64 – 89 %) and information ( 63 – 80 %) while under-using questioning ( 25 – 34 % vs. 42 % for humans). Right: Tactic stickiness. Blue: P(T∈turnt∣T∈turnt−1) ; gray: P(T∈turnt∣T∈/turnt−1) . For humans, whether a tactic appeared in the previous turn has limited influence; for LLMs, the gap is dramatic. The two phenomena compound: tactics that LLMs overuse (e.g., advice ) are also the stickiest, creating a loop of repetitiveness.
Source
# Pairs
Bigram
BLEU-2
BERTSc.
Stick.
Human
265
0.022
0.035
0.865
0.266
GPT-3.5-turbo
1,066
0.100
0.134
0.860
0.516
GPT-4
228
0.089
0.127
0.858
0.556
Llama-2-70B-Chat
103
0.119
0.159
0.860
0.502
Table 1: Cross-turn similarity between consecutive supporter turns. The first three columns measure lexical and semantic overlap; Stick. = tactic stickiness P(T∈turnt∣T∈turnt−1) (the lower the better). Only stickiness reveals that LLMs repeat discourse moves at roughly double the human rate.
Figure 3: Overview of Mint . Step 1 : A multi-turn conversation provides tactic history. Step 2 : The policy πθ generates a new supporter response, and a sentence-level tactic tagger labels each sentence. Step 3 : The tactic distribution of the current turn ( Q ) is compared against the historical profile ( P ) via DKL(Q∥P) for novelty and H(Q) for within-turn breadth, yielding a combined quality-weighted diversity reward optimized with GRPO.
Figure 4: Aggregate empathy vs. tactic stickiness. Mint (Q + D KL ) gives the best trade-off across both model sizes.
Empathy ↑ (1–5 scale)
Neg. Behaviors ↓ (1–5 scale)
Agg. ↑
Tactic Diversity
Model
Family
Method
Words
Valid.
Elab.
Underst.
Unsol. Adv.
Self-Or.
Dismiss.
Agg.
Tac/Turn
Stick. ↓
Human (Gold)
20.6
1.98
1.62
1.47
1.93
1.48
2.29
2.90
1.97
0.27
Vanilla
70.0
3.76
1.55
2.50
1.99
1.07
1.17
3.60
4.49
0.51
Tactic Prompt
66.8
4.09 *
2.22 *
2.93 *
1.60 *
1.16 *
1.07 *
3.90 *
3.73
0.45
Prompt-based
Tactic + History
66.4
3.99 *
2.10 *
2.90 *
1.45 *
1.14
1.04 *
3.89 *
3.63
0.40 *
VS (Vanilla)
42.2
2.84 *
1.52
2.07 *
2.30 *
1.12
1.67 *
3.22 *
3.02
0.41 *
Table 2: Results on the Lend-an-Ear test set ( 315 supporter turns, 50 conversations). Empathy columns report three positive dimensions ( 1 – 5 scale) and three negative behavioral dimensions; Agg. combines positive and reverse-coded negative dimensions. Tac/Turn = unique tactics per response; Stick. = tactic stickiness P(T∈turnt∣T∈turnt−1)↓ . Significance: paired bootstrap ( 10,000 resamples) vs. Vanilla; ∗p<.01 (Bonferroni-corrected across 7 empathy dimensions + stickiness). Gray cells = significantly worse .
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Shared prompt template for the tactic taggers, used for both training and inference. The highlighted placeholder {tactic_definition} is filled per-adapter; {Full_Response} and {Sentence} are filled with the target response and sentence.
Diversity metric
Satisfaction measure
ρ
p
n
Avg. stickiness
Successful
−0.155
0.204
69
Engaged
−0.177
0.146
69
PositiveInteraction
−0.172
0.157
69
UseAgain
-0.287
0.017 *
69
Avg. new tactics
Successful
0.008
0.946
69
Engaged
0.199
0.101
69
Appendix
Table 4: Conversation-level Spearman correlations between tactic diversity and user satisfaction in the 69 SENSE-7 conversations. * p<0.05 .
Figure 6: Prompt for filtering conversations where the user is seeking emotional support (§ 5.1 ).
Valid.
Elab.
Underst.
Unsol. Adv.
Self-Or.
Dismiss.
Mean
Experts 1 & 2
0.25 ∗
0.69 ∗
0.47 ∗
0.70 ∗
0.82 ∗
0.35
0.55
Experts 2 & 3
0.58 ∗
0.73 ∗
0.43
0.69 ∗
0.52 ∗
0.51
0.58
Experts 3 & 1
0.39 ∗
0.62 ∗
0.37
0.54 ∗
0.70 ∗
0.49
0.52
gpt-oss-120b
0.67 ∗
0.54 ∗
0.44 ∗
0.76 ∗
0.71 ∗
0.35 ∗
0.58
Appendix
Table 5: Inter-annotator agreement (weighted Cohen’s κw ) on the 6 Lend-an-Ear ( Kumar et al., 2026 ) empathy evaluation sub-components. Top rows: pairwise agreement among the three expert annotators. Bottom row: gpt-oss-120b evaluated against the expert median, following Kumar et al. (2026) . Valid. = validating emotions, Elab. = encouraging elaboration, Underst. = demonstrating understanding, Unsol. Adv. = unsolicited advice giving, Self-Or. = self-oriented, Dismiss. = dismissing emotions. ∗p<0.01 (Bonferroni-corrected across 6 empathy dimensions).
Figure 7: Prompt template for turn-level empathy evaluation. Placeholders in braces are filled dynamically: framework with the full empathic communication framework (§ G.1 ), few_shot with three expert-scored examples, history / current_turn with the conversation context, and question / rubric with the sub-component criterion (Table 6 ).
Emotion-aware dialogue generation adapts conversational responses to users’ feelings, context, and interaction goals across text, speech, and multimodal settings.
Recent research shows that greater numbers of people are turning to Large Language Models (LLMs) for emotional support, and that people rate LLM responses as more empathic than human-written responses. We suggest a reason for this success: LLMs have learned and consistently deploy a well-liked template for expressing empathy. We develop a taxonomy of 10 empathic language "tactics" that include validating someone's feelings and paraphrasing, and apply this taxonomy to characterize the language that people and LLMs produce when writing empathic responses. Across a set of 2 studies comparing a total of n = 3,265 AI-generated (by six models) and n = 1,290 human-written responses, we find that LLM responses are highly formulaic at a discourse functional level. We discovered a template -- a structured sequence of tactics -- that matches between 83--90% of LLM responses (and 60--83% in a held out sample), and when those are matched, covers 81--92% of the response. By contrast, human-written responses are more diverse. We end with a discussion of implications for the future of AI-generated empathy.
Emma S. Gueorguieva, Hongli Zhan, Jina Suh +4
Department of Psychology, The University of Texas at Austin. · Department of Linguistics, The University of Texas at Austin. · Department of Computer Science and Engineering, The University of Washington. +2
Large language models have demonstrated conversational capabilities, yet empathetic competence remains challenging. Empathetic support is inherently multi-turn and path-dependent: users disclose concerns gradually, emotions evolve over time, and early responses shape trust and receptivity. Reinforcement learning with verifiable emotion rewards provides scalable supervision for long-horizon interactions. However, existing methods evolve the dialogue policy while keeping its training interaction distribution fixed, creating a mismatch between policy competence and training experience. We introduce a dual-loop self-evolution framework driven by verifiable emotion feedback. With the user simulator and verifier frozen, the inner loop optimizes the multi-turn policy using continuous emotion rewards, while the outer loop uses the same outcomes to estimate policy-relative interaction utility and adapt experience. To obtain estimates from sparse, stochastic rollouts, the framework holds the scenario and interaction state constant within each group and prioritizes conditions whose group pass rates lie near the policy's competence boundary. A hierarchical controller shares evidence across support intents, while uncertainty-guided exploration and uniform rehearsal prevent premature exclusion. The resulting distribution generates trajectories, closing both loops without increasing the rollout budget. On SAGE, our framework raises Qwen3-8B Overall from 53.87 to 79.24 and outperforms protocol-matched uniform emotion-reward reinforcement learning by 7.23 points.
Yi Wei, Shuo Jiang, Huaixia Dou +5
Qwen DianJin Team, Alibaba Cloud Computing · Beihang University · School of Computer Science and Technology, Soochow University
Empathy is central to human connection, yet people often struggle to express it effectively. In blinded evaluations, large language models (LLMs) generate responses that are often judged more empathic than human-written ones. Yet when a response is attributed to AI, recipients feel less heard than when comparable responses are attributed to a human. We built a conversation platform in which participants are asked to offer empathic support to an LLM expressing realistic troubles and conducted a randomized experiment collecting 33,938 messages spanning 2,904 text-based conversations between 968 participants and their LLM conversational partners. We find participants report feeling empathy but systematically fail to express it, but an LLM coaching intervention offering personalized feedback on effective empathic communication significantly boosts it without homogenizing participants' responses. Moreover, we derive a data-driven taxonomy of idiomatic empathic expressions in naturalistic dialogues across personal and workplace trouble scenarios. These results advance the scientific understanding of how empathy is expressed and demonstrate a scalable, AI-based intervention for scaffolding and cultivating it.
Aakriti Kumar, Nalin Poungpeth, Diyi Yang +2
Kellogg School of Management, Northwestern University. · Northwestern Institute on Complex Systems, Northwestern University. · Ryan Institute on Complexity, Northwestern University. +3