DD-GEPA: Prompt Optimization for Dialogue Disentanglement Focusing on Task Instruction and Utterance Representation

2026-06-05

AI Summary

This work addresses the challenge of disentangling interleaved conversations in multi-party chats, where topic entanglement often leads to ambiguous and incoherent dialogue structures. To overcome the low accuracy of large language models in conversation disentanglement, the authors propose an automated prompt optimization method that decomposes prompts into three components: task instruction, utterance representation, and output instruction. For the first time, they employ the composite AI system optimization framework GEPA to jointly optimize these components. This approach enables systematic, automatic tuning of critical prompt elements, significantly improving disentanglement accuracy across multiple benchmark datasets. The results not only surpass those achieved with original prompts but also outperform carefully handcrafted ones, demonstrating the effectiveness of structured prompt modeling combined with automated optimization.

Abstract

Multi-party chat often contains interleaved dialogues because multiple participants can discuss different topics at the same time. Dialogue disentanglement addresses this problem by separating an entangled utterance sequence into coherent dialogues. While large language models (LLMs) are promising for this task, they still struggle with dialogue disentanglement and achieve low accuracy. This paper proposes an automatic prompt optimization for LLM based dialogue disentanglement. We decompose the prompt into three components: task instruction, utterance representation, and output instruction, and optimize them using GEPA, an optimization method for compound AI systems. Experiments on benchmark datasets show that the optimized prompts improve dialogue disentanglement accuracy over the original prompts and can surpass hand crafted prompts.

Problem

• dialogue disentanglement
• multi-party chat
• prompt optimization
• large language models
• utterance representation

Innovation

• prompt optimization
• dialogue disentanglement
• large language models
• GEPA
• utterance representation