Discourse Diversity in Multi-Turn Empathic Dialogue

Desmond C. Ong; Emma S. Gueorguieva; Hongli Zhan; Javier Hernandez; Jina Suh; Junyi Jessy Li

arxiv: 2604.11742 · v1 · submitted 2026-04-13 · 💻 cs.CL · cs.AI

Discourse Diversity in Multi-Turn Empathic Dialogue

Hongli Zhan , Emma S. Gueorguieva , Javier Hernandez , Jina Suh , Desmond C. Ong , Junyi Jessy Li This is my paper

Pith reviewed 2026-05-10 14:55 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords empathic dialoguediscourse diversitymulti-turn conversationsreinforcement learninglarge language modelstactic noveltyemotional support

0 comments

The pith

Large language models reuse the same discourse tactics in consecutive turns of empathic dialogue at nearly twice the rate of human supporters.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Large language models produce empathic single-turn replies yet repeat discourse strategies such as questioning or validating from one turn to the next at rates of 0.50 to 0.56, compared with 0.27 for human supporters. This repetition compounds across longer conversations and is missed by ordinary similarity measures. The authors introduce MINT, a reinforcement learning framework that adds a cross-turn tactic novelty signal to an empathy quality reward. On 1.7B and 4B parameter models the combined objective raises aggregate empathy scores by 25.3 percent over standard training and lowers repetition by 26.3 percent on the larger model, beating quality-only and token-level diversity baselines. The work shows that the models already possess empathic capacity but need explicit training to vary their supportive moves as a conversation unfolds.

Core claim

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). 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.

What carries the argument

MINT (Multi-turn Inter-tactic Novelty Training), a reinforcement learning framework that jointly optimizes an empathy quality reward and a cross-turn tactic novelty signal to reduce repetition of discourse moves.

If this is right

LLMs can deliver more varied empathic support across multiple turns when trained with explicit novelty incentives.
Standard similarity metrics fail to detect discourse-level repetition that affects conversation quality.
Joint quality and diversity rewards outperform either signal used alone.
Both smaller and larger models show gains, indicating the approach is not limited to scale.
Effective multi-turn empathy requires planning for strategy variety rather than optimizing each response in isolation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar novelty objectives might apply to other long-horizon dialogue tasks such as tutoring or customer service where repetition reduces engagement.
The results suggest that empathy training corpora alone are insufficient without added diversity constraints.
Future systems could add explicit multi-turn planning modules to select tactic sequences ahead of time.
Longer-term user studies would be needed to confirm whether reduced repetition translates to better real-world outcomes.

Load-bearing premise

Discourse tactics can be identified reliably enough by classifiers to serve as training signals, and rewarding their cross-turn novelty will produce more effective support without harming coherence or other unmeasured qualities.

What would settle it

Human raters find no gain in perceived support quality or engagement for MINT dialogues despite lower measured repetition and higher automated empathy scores, or the tactic classifier shows low agreement with human annotations of strategy use.

Figures

Figures reproduced from arXiv: 2604.11742 by Desmond C. Ong, Emma S. Gueorguieva, Hongli Zhan, Javier Hernandez, Jina Suh, Junyi Jessy Li.

**Figure 2.** 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 p… view at source ↗

**Figure 3.** 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 … view at source ↗

**Figure 4.** Figure 4: Aggregate empathy vs. tactic stickiness. MINT (Q + DKL) gives the best trade-off across both model sizes. MINT Methods. MINT introduces diversity directly into the reward function at the level of discourse moves, rather than regularizing the per-token distribution as in R1-Zero-Div. Building on the quality reward Q established above, we augment it with the cross-turn KL divergence term DKL and the within… view at source ↗

**Figure 5.** Figure 5: Shared prompt template for the tactic taggers, used for both training and inference. [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗

**Figure 6.** Figure 6: Prompt for filtering conversations where the user is seeking emotional support [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗

**Figure 7.** Figure 7: Prompt template for turn-level empathy evaluation. Placeholders in [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗

read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LLMs repeat discourse tactics across turns at double the human rate and MINT RL cuts that while lifting empathy scores, but the tactic labeling step has no validation details.

read the letter

The paper finds that LLMs reuse discourse tactics across consecutive turns at nearly double the rate of human supporters in emotional support conversations. They introduce MINT, a reinforcement learning method that rewards both empathy and cross-turn novelty in tactics to address this. This builds on single-turn findings by looking at how rigidity builds over time in multi-turn settings. The approach combines an empathy reward with a novelty signal based on tactic sequences, and the best version beats vanilla models and other baselines on empathy scores and repetition reduction. It does a good job highlighting why varying discourse moves matters for ongoing support, rather than just sounding empathic in one response. The quantitative improvements on small models are a plus for practical use. The soft spot is the complete lack of information on how the discourse tactics are defined and labeled. The reuse statistic and the MINT novelty term both rely on consistent identification of tactics in supporter responses. If this is done automatically without human validation or if agreement is low, the measured gap between LLMs and humans could be misleading, and the training might optimize for the wrong thing. The abstract gives no details on the taxonomy, annotation process, or any ablations on label quality. It also omits dataset size, how empathy was rated, and whether there are controls for other factors that might affect the results. This is for researchers in conversational AI and emotional support systems who want to improve long-term interaction quality. Readers working on RL for dialogue or discourse analysis might find the framework worth looking at. I would send it to peer review so the methods can be examined in full, particularly the tactic identification process. The core idea seems worth testing properly.

Referee Report

3 major / 2 minor

Summary. The paper claims that LLMs exhibit substantially higher cross-turn reuse of discourse tactics in multi-turn empathic dialogues than humans (0.50-0.56 vs. 0.27), a pattern invisible to standard similarity metrics. It introduces MINT, a reinforcement learning framework that augments an empathy quality reward with a cross-turn tactic novelty signal; the best variant yields a 25.3% gain in aggregate empathy over vanilla LLMs (across 1.7B and 4B models) and a 26.3% reduction in repetition on the 4B model, outperforming quality-only and token-level diversity baselines.

Significance. If the measurement of tactics is reliable, the work identifies a concrete, previously under-examined limitation of LLMs in sustained empathic support and supplies a practical RL recipe that demonstrably increases discourse-move variety while preserving or improving rated empathy. The explicit separation of an external empathy signal from a sequence-level novelty term is a clear methodological strength and supplies falsifiable predictions about multi-turn behavior.

major comments (3)

[§3] §3 (Tactic Taxonomy and Annotation): the headline reuse statistic (0.50-0.56 vs. 0.27) and the MINT novelty reward both presuppose consistent, unbiased labeling of discourse tactics across human and model dialogues. The manuscript reports neither inter-annotator agreement, human validation of the taxonomy, nor an ablation on label noise, leaving both the empirical claim and the 25.3% empathy gain dependent on an unverified measurement step.
[§5] §5 (Results and Evaluation): the reported 25.3% empathy improvement and 26.3% repetition reduction are presented as aggregate figures without accompanying statistical tests, confidence intervals, or controls for conversation length, topic distribution, or model scale. It is therefore impossible to assess whether the gains are robust or could be explained by confounds.
[§4.2] §4.2 (MINT Reward Formulation): the cross-turn novelty term is defined on the same tactic sequences used for the reuse analysis. If tactic labels are produced by an LLM classifier or heuristic that was not independently validated, the RL objective may be optimizing for an artifact of the labeler rather than genuine discourse diversity; no ablation isolating this risk is provided.

minor comments (2)

[§1] The abstract and §1 cite prior work on single-turn repetition (Gueorguieva et al., 2026) but do not clarify how the multi-turn extension differs methodologically from that baseline.
[§4] Notation for the novelty reward (e.g., the precise definition of “cross-turn tactic novelty”) is introduced in §4 but is not restated in the results tables, making it difficult to map numbers back to the objective.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback, which has helped us identify areas where the manuscript can be strengthened. We address each major comment below and have revised the manuscript accordingly to improve clarity, rigor, and completeness.

read point-by-point responses

Referee: [§3] §3 (Tactic Taxonomy and Annotation): the headline reuse statistic (0.50-0.56 vs. 0.27) and the MINT novelty reward both presuppose consistent, unbiased labeling of discourse tactics across human and model dialogues. The manuscript reports neither inter-annotator agreement, human validation of the taxonomy, nor an ablation on label noise, leaving both the empirical claim and the 25.3% empathy gain dependent on an unverified measurement step.

Authors: We agree that explicit reporting of inter-annotator agreement and validation is necessary to support the reliability of the tactic annotations. The taxonomy draws from established frameworks in counseling psychology, and annotations followed detailed guidelines applied by multiple trained annotators. In the revised manuscript we now include Cohen's kappa scores demonstrating substantial agreement, along with a human validation study on a held-out set of dialogues. We have also added an ablation introducing controlled label noise to show that the core reuse statistics and MINT gains remain stable. revision: yes
Referee: [§5] §5 (Results and Evaluation): the reported 25.3% empathy improvement and 26.3% repetition reduction are presented as aggregate figures without accompanying statistical tests, confidence intervals, or controls for conversation length, topic distribution, or model scale. It is therefore impossible to assess whether the gains are robust or could be explained by confounds.

Authors: We acknowledge that the original presentation lacked formal statistical support and confound controls. The revised manuscript now reports paired statistical tests with p-values, bootstrap-derived confidence intervals for the empathy and repetition metrics, and results stratified by conversation length, topic, and model scale. These additions confirm that the reported improvements are statistically significant and consistent across conditions. revision: yes
Referee: [§4.2] §4.2 (MINT Reward Formulation): the cross-turn novelty term is defined on the same tactic sequences used for the reuse analysis. If tactic labels are produced by an LLM classifier or heuristic that was not independently validated, the RL objective may be optimizing for an artifact of the labeler rather than genuine discourse diversity; no ablation isolating this risk is provided.

Authors: We recognize the potential for circularity if the novelty signal were the sole driver. However, the empathy quality reward is computed from an independent scorer trained on human empathy ratings that does not use tactic labels. The joint improvement in both empathy and reduced repetition under MINT supports that the signal captures genuine discourse variation. To isolate the risk, the revision includes an ablation replacing the tactic-based novelty term with an embedding-similarity diversity reward; MINT continues to outperform the baselines on both metrics. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical measurements and RL results are independent of inputs

full rationale

The paper reports new multi-turn empirical statistics on tactic reuse (0.50-0.56 vs. 0.27) and presents MINT as an RL method whose rewards combine an external empathy signal with a novelty term over identified tactics. No equations or derivations reduce the reported gains (25.3% empathy, 26.3% repetition reduction) to parameters fitted on the evaluation data or to the single-turn prior result by construction. The cited Gueorguieva et al. 2026 work is used only for motivation and single-turn background; the central multi-turn findings and training outcomes rest on separate measurements and comparisons to baselines.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claims rest on the ability to annotate and categorize discourse tactics reliably and on the assumption that RL with a combined reward will improve real-world support quality. No explicit free parameters are described in the abstract.

axioms (2)

domain assumption Discourse moves in empathic responses can be consistently categorized into discrete tactics whose repetition can be measured across turns.
The repetition rates and novelty signal depend on this categorization step.
domain assumption Reinforcement learning with an added novelty term will increase tactic diversity without degrading other unmeasured aspects of conversation quality.
Core premise of the MINT training approach.

invented entities (1)

MINT (Multi-turn Inter-tactic Novelty Training) no independent evidence
purpose: Reinforcement learning framework that adds a cross-turn tactic novelty reward to standard empathy optimization.
New method introduced to address the observed repetition problem.

pith-pipeline@v0.9.0 · 5692 in / 1521 out tokens · 76992 ms · 2026-05-10T14:55:12.246228+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

[1]

doi: 10.1609/aaai.v39i13.33569

ISBN 978-1-57735-897-8. doi: 10.1609/aaai.v39i13.33569. URL https://doi.org/10. 1609/aaai.v39i13.33569. Allan Luks and Peggy Payne.The healing power of doing good: The health and spiritual benefits of helping others. iUniverse, 2001. Bethanie Maples, Merve Cerit, Aditya Vishwanath, and Roy Pea. Loneliness and suicide mitigation for students using gpt3-ena...

work page doi:10.1609/aaai.v39i13.33569 2001
[2]

{Tactic}

You will be provided with a full empathic response for context and a single sentence extracted from it. Your task is to determine whether the given sentence contains “{Tactic}”

work page
[3]

{tactic_definition}←varies per adapter

work page
[4]

{Tactic}

Read the sentence and then provide a rating of 0 or 1, with 0 signifying that “ {Tactic}” is not present in the sentence and 1 signifying that “{Tactic}” is present in the sentence. Your response should be in the following format:<score>[]</score>

work page
[5]

{Tactic}

Importantly, the full empathic response is provided as context, and the sentence in question is provided separately. Only the given sentence should be assessed for “{Tactic}”, not the entire response. ###Input: - Context (Full Empathic Response): {Full_Response} - Sentence to Evaluate: {Sentence} ###Response: Figure 5: Shared prompt template for the tacti...

work page 2023

[1] [1]

doi: 10.1609/aaai.v39i13.33569

ISBN 978-1-57735-897-8. doi: 10.1609/aaai.v39i13.33569. URL https://doi.org/10. 1609/aaai.v39i13.33569. Allan Luks and Peggy Payne.The healing power of doing good: The health and spiritual benefits of helping others. iUniverse, 2001. Bethanie Maples, Merve Cerit, Aditya Vishwanath, and Roy Pea. Loneliness and suicide mitigation for students using gpt3-ena...

work page doi:10.1609/aaai.v39i13.33569 2001

[2] [2]

{Tactic}

You will be provided with a full empathic response for context and a single sentence extracted from it. Your task is to determine whether the given sentence contains “{Tactic}”

work page

[3] [3]

{tactic_definition}←varies per adapter

work page

[4] [4]

{Tactic}

Read the sentence and then provide a rating of 0 or 1, with 0 signifying that “ {Tactic}” is not present in the sentence and 1 signifying that “{Tactic}” is present in the sentence. Your response should be in the following format:<score>[]</score>

work page

[5] [5]

{Tactic}

Importantly, the full empathic response is provided as context, and the sentence in question is provided separately. Only the given sentence should be assessed for “{Tactic}”, not the entire response. ###Input: - Context (Full Empathic Response): {Full_Response} - Sentence to Evaluate: {Sentence} ###Response: Figure 5: Shared prompt template for the tacti...

work page 2023