04:00
2026-09-17
machinebrief.com
ai-safety
DyMT-ESB: Dynamic Multi-Turn Evaluation of Social Bias in User-LLM Interactions
A new arXiv paper, DyMT-ESB, introduces a controlled evaluation protocol for social bias in multi-turn user-LLM interactions that generates follow-up user queries from evolving dialogue history rather…