arXiv:2610.28875 records model-specific convergence in LLM-LLM climate attribution debates

The study deploys a repeated-turn debate protocol on climate-change attribution. One agent receives an assigned stance plus optional persuasion instruction; the second receives none. Position is measured by explicit self-report and by turn-level behavioral markers extracted from dialogue. Same-model runs compress the gap on average. Cross-model runs expand it. Explicit persuasion prompts produce no consistent directional lift.

Explicit reports and in-conversation behavior track separate signals. GPT-6 Luna exhibits clearer compression in same-model settings than Gemini 2.5 Flash Lite. Cross-model widening occurs irrespective of which model holds the assigned stance. These patterns hold across ordered pairings, indicating that architecture mismatch itself modulates influence rather than stance assignment alone.

Operational consequence is direct for multi-agent deployments. Systems that mix model families will require separate calibration of opinion drift metrics. Single-model pipelines can rely on gap compression as a stability signal. Neither configuration supports the assumption that an explicit persuade instruction reliably increases influence.

Next measurement step is extension to three-model chains and to topics outside climate attribution to test whether the same-model narrowing and cross-model widening generalize beyond the four combinations tested.