One example:
Gemini will get into modes where it strongly and illogically agrees with whatever it said previously. It feels like sycophancy but isn't purely with whatever the user wants, or perhaps it is modeling what the user, or rather, the unnamed RL reward determinant agent, wants on a more meta-level.
Here, using multiple models is an obvious improvement since (e.g. o3) despite possible being less capable, is less prone to getting into "always defend idea / approach XYZ" basin.
If you use two Geminis, the arbiter will simply get stuck in the opposite opinion basin as the submitter, and then they will argue with each other forever without getting anywhere.
Another pattern is multiple Geminis in different basins (basically steelman agents of different arguments) + one OpenAI model arbiter
Gemini will get into modes where it strongly and illogically agrees with whatever it said previously. It feels like sycophancy but isn't purely with whatever the user wants, or perhaps it is modeling what the user, or rather, the unnamed RL reward determinant agent, wants on a more meta-level.
Here, using multiple models is an obvious improvement since (e.g. o3) despite possible being less capable, is less prone to getting into "always defend idea / approach XYZ" basin.
If you use two Geminis, the arbiter will simply get stuck in the opposite opinion basin as the submitter, and then they will argue with each other forever without getting anywhere.
Another pattern is multiple Geminis in different basins (basically steelman agents of different arguments) + one OpenAI model arbiter