Scientists Find ‘Digital Personalities’ Significantly Improve AI Agents’ Reasoning Accuracy on Complex Tasks

Multi-agent AI systems typically have participants speak in a fixed rotation. While easier to control, that format struggles to reproduce the interruptions, silences and mutual corrections of human discussion. Researchers at the University of Electro-Communications and Japan’s National Institute of Advanced Industrial Science and Technology therefore incorporated psychology’s Big Five personality traits into LLMs, allowing agents to decide when to speak based on their personalities and the urgency of the situation. The work has implications for the reliability of collaborative AI reasoning and group decision-making.

The team released the research on February 5, 2026, and tested three dialogue modes using 1,000 MMLU questions. When one agent initially gave the wrong answer, accuracy was 68.7% with fixed turn-taking, 73.8% with dynamic turn-taking and 79.2% when interruptions were allowed. When two agents initially answered incorrectly, the respective rates were 37.2%, 43.7% and 49.5%, showing that the interruption mechanism also led in more difficult scenarios.