AuraTracer智迹闻
中文

EVENT DOSSIER

Controlling and Assessing Appropriate Persona Use in LLM-based Dialogue Generation

2026-09-07 12:00 Models 🔥 40.2 heat score
1sources
1days unfolding
40.2heat score
3mentions
SummaryAI generated

A paper published on September 7, 2026, on arXiv discusses how to control and evaluate appropriate character settings in dialogue generation based on large language models. This study aims to address the issues of confusion in characters or inappropriate character portrayals that may occur in current generated content, and proposes corresponding control mechanisms and evaluation criteria to optimize the quality and ethical compliance of dialogue generation.

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
LLMPASSCONPOS

Coverage · reports per dayLANGUAGE SPLIT

Entity relations
LLM × PAS1LLM × SCONPOS1PAS × SCONPOS1

SignalsSIGNALS

Keyword heat
  • LLM1
  • SCONPOS1
  • PAS1

All reports (1)SOURCES

A arXiv cs.CL en 2026-09-07 12:00

Controlling and Assessing Appropriate Persona Use in LLM-based Dialogue Generation

针对大语言模型(LLM)在基于角色的对话生成中过度使用角色属性的问题,研究者提出自我对比角色过度抑制方法(SCONPOS)及角色适用性评分指标(PAS)。分析显示,现有 LLM 存在系统性偏差以纳入所有给定角色属性,且现有指标无法捕捉语境适用性。SCONPOS 通过在提示编码阶段直接干预 LLM 内部表示来抑制过度使用,无需响应生成;PAS 则对过度使用和不足使用进行惩罚。实验结果表明,SCONPOS 系统性地减少了过度使用,PAS 有效捕捉了角色使用的语境适用性。