Yin Min Pa Pa | Search arXiv

arXiv2026

APO-DD: Dialogue Disentanglement via Reflective Optimization of Multi-Component Prompts

In multi-party chat, interleaved dialogues make it difficult to identify the context of each utterance. Dialogue disentanglement separates these dialogues to support dialogue analysis and response generation. Prompt refinement can improve large language models' (LLMs') accuracy on this task, but manually finding effective prompts for each LLM requires extensive trial and error. We propose APO-DD, which adapts GEPA to optimize prompts using execution traces and reflections on failed decisions. It updates task instructions directly and chat representations through automatically generated formatting functions. We compare optimizing either component alone or both. With Qwen3.5-9B, the best optimized prompt achieves an average score of 68.7 across seven evaluation metrics on Ubuntu IRC, compared with 57.5 for the best manually refined prompt. It also outperforms manually refined prompts on other chat domains without further optimization. For this model, optimizing task instructions alone does not always improve accuracy over the initial prompts. Our results suggest that exploring chat representation as an optimization target may yield higher accuracy than optimizing task instructions alone. Our code is available at https://anonymous.4open.science/r/DB6B.

cs.SE↗

arXiv2026

Understanding Interpretation Difficulty in Harmful Online Communication: Insights from Cybercrime Communities

Harmful online communication often contains slang, coded terms, abbreviations, and community-specific expressions, which make messages difficult to interpret. This paper presents an exploratory study of interpretation difficulty in Discord chats related to cybercrime. We construct reference interpretations of purposefully selected difficult messages, which were reviewed by an expert. We then use them to evaluate human and large language model (LLM) interpretations under different context conditions. The results show that local context alone is often insufficient for humans, while external knowledge and extended conversational context substantially improve human interpretation. For LLMs, local context also improves interpretation, and the larger model performs better. We further conduct a qualitative error analysis and propose a preliminary classification of factors that make harmful chats difficult to interpret. These findings suggest that harmful-content analysis should treat interpretation as an evidence-integration problem, rather than as message-level classification alone.

cs.CL↗