Following the TRACE: A Structured Path to Empathetic Response Generation with Multi-Agent Models

Empathetic response generation is a crucial task for creating more human-like and supportive conversational agents. However, existing methods face a core trade-off between the analytical depth of specialized models and the generative fluency of Large Language Models (LLMs). To address this, we propose TRACE, Task-decomposed Reasoning for Affective Communication and Empathy, a novel framework that models empathy as a structured cognitive process by decomposing the task into a pipeline for analysis and synthesis. By building a comprehensive understanding before generation, TRACE unites deep analysis with expressive generation. Experimental results show that our framework significantly outperforms strong baselines in both automatic and LLM-based evaluations, confirming that our structured decomposition is a promising paradigm for creating more capable and interpretable empathetic agents. Our code is available at https://anonymous.4open.science/r/TRACE-18EF/README.md.

💡 Research Summary

The paper addresses a central dilemma in empathetic response generation: specialized models provide deep analytical understanding but lack linguistic fluency, while large language models (LLMs) generate fluent text but often miss nuanced emotional insight. To bridge this gap, the authors introduce TRACE (Task‑decomposed Reasoning for Affective Communication and Empathy), a multi‑agent framework that mirrors the cognitive stages of human empathy. TRACE decomposes the task into four sequential modules, each handled by a dedicated agent:

1. Affective State Identifier (ASI) – Takes the dialogue history and predicts a core emotion label. It maps 32 fine‑grained emotions to Ekman’s six basic categories, using a simple arg‑max probability formulation to ensure robust, interpretable emotion grounding.

2. Causal Analysis Engine (CAE) – Given the identified emotion, CAE extracts local trigger spans, generates a global cause summary, and assigns a psychological cause category from a predefined taxonomy. This dual‑granularity analysis yields a structured object (trigger spans, summary, category) that captures why the user feels the way they do.

3. Strategic Response Planner (SRP) – Selects a communicative strategy (e.g., Emotional Reaction, Interpretation, Exploration) from a small predefined set. SRP leverages a Retrieval‑Augmented Generation (RAG) subsystem to fetch similar conversation scenarios from the training corpus, measuring semantic similarity with cosine similarity of embedding vectors. If exact matches (emotion‑aligned and similarity‑above a threshold τ) are unavailable, a fuzzy search relaxes the constraints. The retrieved exemplars inform a strategy‑selection function that outputs the optimal strategy s*.

4. Empathetic Response Synthesizer (ERS) – The final agent synthesizes the emotion label, causal analysis, and chosen strategy into a prompt for GPT‑4o, again using RAG to retrieve stylistic exemplars. This ensures the generated reply not only reflects the analytical context but also mimics human‑like tone and phrasing.

Experiments are conducted on the widely used ED (EmotionLines) dataset. Baselines span three categories: (i) specialized models (Multi‑TRS, EmpDG, KEMP, CEM, CASE, EmpSOA), (ii) pre‑trained dialogue models (BlenderBot, DialoGPT, LEMPEx), and (iii) LLM‑based methods (EmpGPT‑3, EmpCRL). Automatic metrics include Perplexity (fluency), Distinct‑n and EAD‑n (lexical diversity), and I‑ACC (emotion accuracy). TRACE achieves state‑of‑the‑art results across the board, notably attaining Distinct‑1 13.62, Distinct‑2 48.12, and I‑ACC 44.28, far surpassing all baselines.

For human‑like evaluation, the authors employ GPT‑4o as an automated assessor, performing pairwise A/B tests on 100 random samples. TRACE wins over the raw GPT‑4o backbone in empathy (80% vs. 20%), informativeness (74% vs. 26%), fluency (79% vs. 21%), and consistency (85% vs. 15%). Compared with EmpGPT‑3, TRACE also shows significant gains on all criteria.

Ablation studies remove each component (RAG, ASI, CAE, SRP) in turn. The removal of the full analysis pipeline or RAG causes the steepest drops in diversity scores, confirming that both structured reasoning and exemplar retrieval are critical. Smaller degradations from omitting ASI, CAE, or SRP indicate that each analytical layer contributes incrementally to the richness of the final response.

In conclusion, TRACE demonstrates that explicitly modeling empathy as a structured pipeline of analysis, planning, and generation can simultaneously achieve deep emotional understanding and high‑quality, diverse language output. The framework’s modularity offers interpretability, facilitates integration of external knowledge, and paves the way for more trustworthy, empathetic conversational agents in domains such as mental‑health support, customer service, and human‑robot interaction.