Research map / Large language models / Language, safety and interpretability

SocialSim: Towards Socialized Simulation of Emotional Support Conversation

AAAI 2025 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SocialSim, a framework that simulates emotional support conversations via seeker persona realism and supporter cognitive reasoning, and uses it to generate the SSConv corpus (3,229 dialogues) for training a Llama-2-7b chatbot that outperforms ESConv- and ExTES-trained models.

Problem

Synthetic emotional support conversation (ESC) data generated by LLMs overlooks the social dynamics of ESC, so seeker scenarios lack specificity and supporter responses rely on surface-level patterns rather than cognitive reasoning, leaving a gap with crowdsourced corpora.

Approach

Seeker-side persona realism builds a bank of structured personas from real PsyQA help-seeking scenarios to elicit social disclosure, and supporter-side chain-of-thought nodes (Situation, Thought, Action, Strategy) elicit cognitive reasoning before each response to enhance social awareness.

Key results

• On SSConv-test, SSConv-trained Llama-2-7b with reasoning (SSConv•) reaches BLEU-1 32.19 and NAvg 1.390 vs ESConv-trained 1.000 and ExTES-trained 1.198.
• In interactive human evaluation (216 sessions), SSConv◦ vs ESConv◦ overall win/loss/tie was 65/2/5.
• Human quality ratings for SSConv (2.76 informativeness, 3.00 safety) exceed ESConv and ExTES on most criteria; on ESConv-test all models perform similarly.

Measured effects reported in the paper

Surprising findings

• On the in-domain ESConv-test set, models trained on synthetic data perform on par with the ESConv-trained model, with no significant differences.
• The 65,000-dialogue AugESC baseline underperforms the 3,229-dialogue SSConv, indicating data quality matters more than quantity.

Limitations

• Persona bank and dialogues are derived from a Chinese PsyQA dataset translated to English with GPT-4, and the synthetic data relies on GPT-4 generation.
• Automatic metrics measure semantic overlap with golden responses and do not fully capture performance, as the authors note.
• Human evaluation uses a small number of workers and sessions (30 workers, 216 sessions).

Experimental setup

Datasets: PsyQA ESConv AugESC ExTES EmpatheticDialog

Benchmarks: SSConv-test ESConv-test

Metrics: BLEU-1 BLEU-2 ROUGE-L METEOR Extrema Distinct-1 Distinct-2 NAvg human quality scores (informativeness, understanding, helpfu interactive human win/loss/tie

Baselines: ESConv ExTES AugESC

Builds on: chain-of-thought prompting ESConv PsyQA Hill helping skills theory

Closest papers by meaning

• OdysSim: Building Foundation Models for Human Behavior Simulation (NeurIPS 2026 · similarity 0.87)
• Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.86)
• MimeQA: Towards Socially-Intelligent Nonverbal Foundation Models (NeurIPS 2025 · similarity 0.85)
• RLVER: Reinforcement Learning with Verifiable Emotion Rewards for Empathetic Agents (ICLR 2026 · similarity 0.85)
• An LLM-based Simulation Framework for Embodied Conversational Agents in Psychological Counseling (AAAI 2026 · similarity 0.85)
• EmpathyChat: Structured Cognitive Reasoning in Empathetic Spoken Dialogue (NeurIPS 2026 · similarity 0.84)