LLM-LLM debate study shows model pairing and measurement shape opinion change
The paper (arXiv:2610.28875v1) introduces a controlled LLM‑LLM debate framework where two models discuss climate‑change attribution under varied roles and instructions. Across four ordered pairings of GPT‑6 Luna and Gemini 2.5 Flash Lite, the authors report that same‑model interactions tend to narrow opinion gaps (but clearly for only one model), cross‑model interactions tend to widen them, explicit persuasion instructions do not reliably increase influence, and self‑reports versus in‑dialogue behavior capture different aspects of opinion dynamics.