Researchers Develop Methods to Improve Role Consistency and Emotional Expression in AI Agent Systems

Researchers Develop Methods to Improve Role Consistency and Emotional Expression in AI Agent Systems

Researchers have published new approaches to address critical challenges in AI agent systems, focusing on role consistency in multi-agent collaboration and emotional expression in speech synthesis.

According to arxiv.org, a paper on improving role consistency introduces a quantitative method to address “disobey role specification,” where agents fail to adhere to their assigned responsibilities in large language model (LLM)-driven multi-agent systems. The researchers propose using a role clarity matrix to measure alignment between agents’ role descriptions and their behavioral trajectories. When tested on the ChatDev multi-agent system, the method reduced role overstepping rates dramatically: with Qwen, the rate decreased from 46.4% to 8.4%, and with Llama, from 43.4% to 0.2%. The role clarity score increased from 0.5328 to 0.9097 for Qwen and from 0.5007 to 0.8530 for Llama, according to the paper.

In a separate development, arxiv.org reports on WiseMind, a multi-agent framework for psychiatric assessment that combines a “Reasonable Mind” agent for evidence-based logic with an “Emotional Mind” agent for empathetic communication. Using a DSM-5-guided knowledge graph, WiseMind achieved 85.6% top-1 diagnostic accuracy across 1,206 simulated conversations and 180 real user sessions, surpassing knowledge-enhanced single-agent baselines by 15-54 percentage points.

Additionally, arxiv.org describes a two-stage prompt selection strategy for expressive speech synthesis that evaluates prompts using pitch-based features, audio quality, and emotional alignment to improve both emotional intensity and speaker identity stability in zero-shot text-to-speech systems.