SyncAI.news, a Varaisys broadcasting
EmoTrace: Recursive Trace-Factorized Generation for Emotional Support Conversation
MZ

Mingwen Zhang, Minqiang Yang, Yang Yu, Haobo Liang, Yongfeng Tao, Changsheng Ma, Bin Hu

· 1 min read

ResearcharXiv cs.CL

EmoTrace: Recursive Trace-Factorized Generation for Emotional Support Conversation

arXiv:2611.18628v1 Announce Type: new Abstract: Fundamental to effective emotional support is the iterative interpersonal dynamic wherein a supporter intervention continuously shapes the psychological trajectory of the seeker. While contemporary models capture variables at the turn level, including transient emotions and discrete strategies, they inherently treat these factors as isolated events. By retaining strictly literal dialogue histories while discarding the underlying decision logic, such architectures remain fundamentally incapable of sustaining coherent extended support. Here we introduce EmoTrace, a recursive method utilizing trace factorization to preserve explicit decision continuity throughout interactions spanning multiple turns. Departing from others, our approach formulates each conversational turn as a structured interaction increment encapsulating the evolving state of the seeker, the selected strategy, the affect of the supporter, and the ultimate textual realization. Accumulated sequentially, this trajectory of increments functions as an explicit structural trace informing subsequent interventions. Optimization of this factorized generation is achieved through two complementary mechanisms. Specifically, a value guidance mechanism aligns discrete factor selection with interaction values over an extended horizon, whereas a trace conditioning mechanism refines textual realization contingent upon the fixed strategic trace. Empirical evaluations across standard benchmarks demonstrate that EmoTrace yields substantial improvements in both strategic planning and response generation. Crucially, rollouts spanning multiple turns reveal that these retained structural traces confer compounding benefits over extended dialogues, establishing a robust mechanism for sustained affective alignment in human-machine interaction.

Original source

This story was published by arXiv cs.CL and written by Mingwen Zhang, Minqiang Yang, Yang Yu, Haobo Liang, Yongfeng Tao, Changsheng Ma, Bin Hu. SyncAI.news shows a preview; the complete article is on the publisher's site.

Read the full story on arxiv.org

Similar News