PaperlayerView exact revision on arXiv

Source-linked AI summary

ESC-Skills: Discovering and Self-Evolving Skills for Emotional Support Conversations

Jie Zhu, Huaixia Dou, Shuo Jiang, Junhui Li, Lifan Guo, Feng Chen, Chi Zhang, Fang Kong

arXiv:2605.27908v1cs.CLcs.AI

TL;DR

ESC systems have limited interpretability and systematic skill improvement because intervention effects on seekers’ emotional states are not explicitly represented or verified. ESC-Skills models these effects with Intervention Units, builds and refines an executable Skills Bank through multi-profile simulation, and improves response quality and long-horizon emotional support outcomes across LLM backbones. The evaluation remains simulation-based rather than based on live user studies.

  • Problem

    ESC research has underexplored how interventions affect subsequent seeker emotional states and how intervention knowledge can be explicitly represented, verified, and continually improved.

  • Method

    ESC-Skills represents localized support interactions as Intervention Units, constructs an executable Skills Bank from successful and failed dialogues, and refines it through multi-profile SAGE simulation and verification.

  • Results

    ESC-Skills consistently improves response-level quality and long-horizon emotional support outcomes across multiple LLM backbones on ESConv and SAGE.

  • Takeaways & Limitations

    The framework provides executable, state-aware support skills with more interpretable and controllable intervention behavior.

  • Takeaways & Limitations

    Evaluation uses SAGE simulated help-seekers rather than live user studies, so it does not capture the full variability of real human emotional responses.

Abstract

from arXiv · show

Existing emotional support conversation (ESC) systems mainly rely on end-to-end response generation or coarse strategy supervision, offering limited interpretability and little support for systematic skill improvement. We propose ESC-Skills, a skill-centric framework that discovers and self-evolves executable emotional support skills. We first model localized support interactions as Intervention Units (IUs), which capture state--action--outcome dynamics between seeker states, support interventions, and post-response emotional changes. Based on IUs extracted from both successful and failed ESC dialogues, we construct the ESC-Skills Bank, a repository of executable emotional support skills containing intervention guidance, applicability conditions, expected outcomes, and potential risks. To further improve robustness, we introduce a multi-profile self-evolutionary refinement framework in which an ESC agent interacts with diverse simulated seeker profiles under SAGE evaluation. The resulting interaction traces are analyzed to identify missing skills, unsafe interventions, and profile-specific failure patterns, which are then used to refine the Skills Bank through simulation-based verification. Experimental results demonstrate that ESC-Skills improves both response-level quality and dialogue-level emotional outcomes while providing more interpretable and controllable support behaviors. We will release the code, prompts, and ESC-Skills Bank at https://github.com/aliyun/qwen-dianjin.

1 Introduction

ESC-Skills reframes emotional support as intervention-driven state transitions, addressing the need to represent, verify, and improve how interventions affect seekers’ subsequent emotional states. It combines executable skill construction with multi-profile refinement to make support behaviors more suitable and controllable.

  • Motivation: Existing ESC research has improved response generation and support-strategy control, but intervention effects on subsequent seeker emotional states remain underexplored.The gap includes how intervention knowledge should be explicitly represented, verified, and continually improved.
  • Motivation: Effective ESC requires selecting interventions that induce beneficial emotional state transitions, not only generating empathetic responses.A generic suggestion can overlook underlying self-doubt and fear, whereas a suitable intervention can facilitate self-awareness and constructive emotional reflection.
  • Approach: ESC-Skills models localized support interactions as Intervention Units capturing seeker states, support actions, and post-response emotional changes.This skill-centric formulation treats emotional support as localized state–action–outcome intervention dynamics.
  • Approach: The ESC-Skills Bank stores executable skills induced from successful and failed dialogues, including applicability conditions, intervention guidance, expected outcomes, risks, and failure-prone anti-patterns.Using both successful and failed interactions supports representation of effective patterns alongside unsafe or failure-prone ones.
  • Approach: A multi-profile self-evolutionary refinement framework uses simulated interactions and verification to identify missing skills, unsafe interventions, and profile-specific failures.The framework enables continual skill refinement through simulation-based verification.

2 Related Work

Prior ESC work emphasizes strategy selection, planning, synthetic data, reasoning, and collaboration, while general self-improving agent methods target domains with clearer success signals. ESC-Skills addresses this mismatch by representing affect-grounded intervention knowledge and evaluating it through simulated interaction signals.

  • Emotional support conversations: ESC research has followed a strategy-predict-then-generate paradigm using commonsense, state-transition planning, synthesized dialogues, chain-of-thought reasoning, and multi-agent collaboration.These approaches primarily target strategy selection, response generation, or dialogue planning.
  • Self-improving agent skills: Self-improving agent research uses recursive reinforcement learning, sandboxed optimization, self-evolutionary verification, reflective memory, and lifecycle governance.SkillsBench identifies closed-loop feedback as critical for effective skill improvement.
  • Research gap: Existing self-improvement methods often assume relatively clear success signals, whereas emotional support conversations lack a reliable deterministic oracle.This difference makes direct transfer of conventional skill-improvement approaches difficult.
  • Research gap: ESC-Skills represents expertise as structured SKILL.md packages and evaluates it through simulation-based interaction signals grounded in the seeker’s affective state.The representation targets behavioral intervention knowledge rather than generic executable code, tool procedures, or prompt heuristics.

3 Methodology

ESC-Skills models emotional support as localized intervention dynamics, builds executable skills from successful and failed dialogues, and refines them through multi-profile simulation-based verification.

  • 3.1 Problem Definition: Intervention Units represent each localized interaction as a seeker state, support action, and resulting post-intervention emotional state.Resulting transitions may be constructive, such as relief or openness, or negative, such as withdrawal or increased distress.
  • 3.2 ESC-Skills Bank Construction: The extraction process annotates scenarios, fine-grained seeker states, support actions, and post-response emotional changes across successful and failed ESC dialogues.The data include 910 ESConv conversations and 196 FailedESConv conversations, with 17 action types and 15 seeker states.
  • 3.2 ESC-Skills Bank Construction: 17,858 Intervention Units were extracted, including 10,181 key units with 9,697 positive and 484 negative emotional-shift instances.Key units capture salient positive or negative post-intervention changes, while weak or stable changes are treated as non-key.
  • 3.2 ESC-Skills Bank Construction: The 258 prototype groups are clustered by semantic similarity and synthesized into executable SKILL.md documents containing activation conditions, recommended actions, and pitfalls.The initial ESC-Skills Bank contains 27 executable emotional support skills, denoted B0.
  • 3.3 Multi-Profile Self-Evolutionary Skill Refinement: The refinement framework simulates interactions with 500 seeker profiles, analyzes profile-specific performance, and verifies refined or newly generated skills on 15 challenging profiles.Updates are accepted after universal success or strict improvement in average emotion score within three attempts; otherwise they are rolled back or removed, producing B⋆ with 34 skills.

4 Experimentation

Experiments evaluate ESC-Skills on response-level ESConv performance, dialogue-level SAGE outcomes, baseline comparisons, and component ablations. Results show consistent gains from refined skills, while static or unverified skills are less robust.

  • Experimental Setup: ESC-Skills is evaluated on ESConv’s 195-conversation test split and SAGE’s 100 predefined seeker profiles using response- and dialogue-level metrics.Response metrics cover strategy accuracy and response quality; SAGE measures average sentient score, successful dialogues, and failures.
  • Main Results: Across multiple LLM backbones, ESC-Skills improves ESConv strategy accuracy and generally improves BLEU, ROUGE, METEOR, and BERTScore.Qwen3.6-Plus gains 12.06% accuracy, while Claude-Haiku-4.5’s BERTScore rises from 69.13 to 84.03.
  • Main Results: On SAGE, Qwen3.6-Plus rises from 66.4 to 72.1 average sentient score and from 13 to 31 successful dialogues after skill augmentation.Similar improvements are reported for Gemini-3.1-Flash and Claude-Sonnet-4.6, alongside fewer severe failures.
  • Comparison to Baselines: For Qwen3.6-Plus, ESC-Skills outperforms alternative skill baselines across response- and dialogue-level evaluation, including ESConv ACC improving from 11.50 to 23.56.Self-generated, chain-of-thought-guided, SkillCreator, and human-curated skills provide marginal or inconsistent gains.
  • Ablation Study: The initial static bank improves some ESConv metrics but lowers SAGE average emotion scores from 66.4 to 61.1 and increases failures from 14 to 19.Interaction-driven updates without verification partially recover performance but remain limited over the no-skill baseline.
  • Ablation Study: The final refined bank achieves the best overall performance across both benchmarks, with verification filtering ineffective skills and improving intervention robustness.On SAGE, it more than doubles successful dialogues relative to no skills and improves ESConv ACC and METEOR.
  • Human Evaluation: Human evaluation is largely consistent with automatic and GPT-Judge metrics, with Claude-Haiku-4.5 showing the largest human Overall gain of ∆Overall = +0.44.Three annotators rated Empathy, Helpfulness, and Overall quality on 100 ESConv test instances.

5 Conclusion

ESC-Skills models emotional support as an intervention-driven process and builds executable skills from localized interaction dynamics. Across ESConv and SAGE, the framework improves response quality and long-horizon emotional outcomes across multiple LLM backbones.

  • Contributions: ESC-Skills represents emotional support with Intervention Units capturing localized seeker-state, intervention, and post-response outcome dynamics.The framework constructs an executable Skills Bank from successful and failed dialogues, then refines it through multi-profile SAGE verification.
  • Findings: Experiments on ESConv and SAGE show consistent improvements in response-level quality and long-horizon emotional support outcomes across multiple LLM backbones.The findings support skill-centric and interaction-driven approaches for building more robust emotional support agents.

Limitations

The evaluation is limited to simulated help-seekers and English-language counseling, while expert review and continuous online evolution remain future work.

  • Evaluation scope: SAGE enables controlled, large-scale comparisons but does not capture the full variability of real human emotional responses.Complementary evaluation with trained counselors is identified as a natural next step.
  • Domain and language coverage: The current ESC-Skills instantiation targets English-language emotional support counseling and has not been validated in other supportive dialogue settings or languages.The evolution framework is described as domain-agnostic, with broader validation planned.
  • Skill review: The current pipeline lacks human expert review, so licensed counselor oversight would add a safety layer in clinical or high-risk deployments.This oversight is proposed within the evolution loop.
  • Base model requirements: ESC-Skills is demonstrated on strong instruction-following LLMs, leaving transfer to smaller or open-weight models and capacity-adaptive skill complexity unresolved.These questions remain open directions.
  • Online evolution: The deployed Skills Bank is fixed at deployment time rather than continuously updated from live interaction signals without regression.Safe online evolution is left as a non-trivial extension for future work.

Ethics Statement

The study uses anonymized public peer-support data and third-party LLM APIs, with safeguards for human annotators and reproducibility materials for evaluation.

  • Data: The study uses the publicly released ESConv corpus, which contains anonymized peer-support dialogues and no additional personally identifiable information.The corpus is used under its original research-use license.
  • Intended use and risks: ESC-Skills is a research artifact, not a substitute for licensed professionals, and crisis or clinical deployment requires expert oversight, safety auditing, classifiers, and human escalation.Generated responses may occasionally fail to recognize crisis signals.
  • Human annotation: Three English-proficient annotators received sensitive-content warnings, opt-out options, mental-health resources, and above-minimum-wage compensation.Detailed evaluation guidelines are provided in the appendix.
  • LLM usage: The agent and judge rely on third-party LLM APIs; model weights are not released, while the skill bank and evaluation code will be public.API calls comply with provider terms of service.
  • Reproducibility: Prompts, evolution hyperparameters, sampled turn indices, and aggregated judge scores will be released to support reproduction without redisclosing raw seeker utterances.The raw utterances are limited to what is already public in ESConv.
  • Annotation and intervention units: Intervention Units are built from structured annotations of scenarios, seeker states, supporter actions, and post-intervention response changes.The label sets were developed through ESConv and FailedESConv inspection plus preliminary LLM-based open coding.

B More Details of the Skill Prototypes

Skill prototypes quantify how intervention patterns affect post-response emotions, while executable skill documents organize these patterns into a deployable bank.

  • Prototype effectiveness: Eight representative skill prototypes have a 100.0% effectiveness rate, with all associated IUs producing positive post-response emotional changes.These prototypes suggest that some support actions work consistently under specific seeker states.
  • Prototype filtering: Below-threshold prototypes show that identical support actions applied to identical seeker states can produce substantially different emotional outcomes across conversations.Some interventions can cause agitation, withdrawal, or confusion in particular contexts.
  • Skill artifacts: The esc-strategy-switching skill is a 461-line, eight-section SKILL.md example, accompanied by esc-action-planning with the same structure.The excerpt preserves metadata, headings, and representative bullets while replacing remaining content.
  • Final skill bank: B⋆ contains 34 skills, each marked as inherited, updated, or newly introduced during evolution.The skills are grouped by metadata.category.

E Case Study: A Strategy-Switching Failure Mode

A case study shows five systems repeating an empathy-first response when the seeker is ready for action, whereas B⋆ switches to concrete suggestions matching the gold strategy.

  • Case setup: The case compares six support arms on one ESConv supporter turn, with the complete comparison reported in Table 12.The example is used to make aggregate gains concrete.
  • Dialogue context: The seeker’s confirmation request follows two suggestion turns and signals readiness for a next concrete action.The gold response is consequently labeled PROVIDING SUGGESTIONS.
  • Baseline behavior: Five arms select REFLECTION OF FEELINGS and produce near-paraphrastic empathy-first openings on this turn.QWEN3.6-PLUS, +SELF-GEN, +COT-SELF-GEN, +SKILLCREATOR, and +HUMANCURATED converge on the same response tendency.
  • Strategy switching: +ESC-SKILLS is the only arm selecting PROVIDING SUGGESTIONS, matching the gold annotation with specific actions such as updating a resume or browsing one job board.esc-strategy-switching detects the confirmation cue, then esc-action-planning produces the concrete plan; both skills arose from repeated over-empathy failures.
  • Quantitative connection: +ESC-SKILLS contributes +8.7 over Base on the main-table SAGE delta, while the five non-Ours replies show deltas of ≤1.5 points.The case localizes the practical gap to switching from empathy to advice when the seeker signals readiness.
  • Takeaway: The case illustrates the simplest strategic transition, empathy to action, which only the SAGE-evolved bank handles correctly.The authors distinguish this coverage from fluent skill generation alone.

F Detailed Performance

This section presents detailed evaluations across response-level and dialogue-level settings, alongside skill-bank composition and prototype filtering. It also documents the prompt templates used to support reproducibility.

  • Dialogue-level evaluation: Table 14 reports dialogue-level SAGE metrics including emotion scores, success and failure counts, and final emotional-grade distributions.Higher Avg.Score, Media, Max, and Success indicate better support quality, while lower Failures indicates fewer harmful interactions.
  • Reproducibility: The appendix lists verbatim prompt templates underlying the key LLM calls in ESC-Skills to make the pipeline reproducible.
  • Prototype selection: Eight skill prototypes achieved an effectiveness rate of 1.0, while other prototypes were filtered out for falling below the effectiveness threshold.

H Human Evaluation Annotation Guidelines

The evaluation guidelines specify how human annotators rate ESC responses, while the surrounding materials document the skill-bank prompts, strategy definitions, evolution procedures, and comparison settings. Together, they define the evaluation and implementation context for skill-based emotional support.

  • Human evaluation: Human evaluators rate supporter responses on empathy, helpfulness, and overall quality using a 1–5 Likert scale.Three independent annotators evaluate sampled ESConv supporter turns, with model identities hidden and response order randomized to reduce annotation bias.
  • Skill representation: The skill bank includes executable skills for intervention strategy switching, with triggers for changed seeker states and repeated unmet advice-seeking bids.
  • Evaluation settings: Evaluation materials compare response-level and dialogue-level outcomes across skill-based baselines and SAGE benchmark settings.
  • Skill-based evaluation: The ESC agent prompt inserts the full skill-bank content or leaves the skills section empty for a no-skill baseline.
  • Skill evolution: Skill evolution analyzes evaluation conversations, then updates existing skills or creates new ones based on aggregated recommendations.Update prompts preserve the SKILL.md YAML frontmatter, while creation prompts use low-scoring conversations and the existing catalogue to avoid overlap.
  • Human evaluation: The task presents the seeker’s situation, dialogue history, and one supporter response for evaluation.
Loading 2605.27908v1…