We introduce a controlled LLM‑LLM debate framework where two language models repeatedly discuss a contested issue—climate‑change attribution.
One model receives an assigned stance and, in certain conditions, an explicit instruction to persuade; the other model receives no preset position and forms its view solely from the dialogue.
Four ordered model pairings are tested: same‑model pairs of GPT‑6 Luna and Gemini 2.5 Flash Lite, as well as cross‑model pairs. For each turn we log both the role’s explicitly reported position and a separate turn‑by‑turn behavioral measure.
The findings reveal that same‑model interactions tend to narrow the opinion gap on average, but this effect is clear only for GPT‑6 Luna. Cross‑model interactions, by contrast, widen the gap. Explicit reports and in‑dialogue behavior capture distinct aspects of opinion dynamics.
Importantly, a blanket “persuade” instruction does not reliably boost influence, indicating that persuasive power depends on the intrinsic characteristics of the models and their counterpart.
These results suggest that conversational influence in extended LLM interactions is shaped not just by assigned roles and initial positions, but also by the specific models involved, raising new considerations for measuring and guiding opinion change.
Review