A recent study[1] seems to indicate that polarization is a hard problem, along with some of the other negative effects of social media. Many of the commonly suggested solutions have minimal impact, or no effect at all. That flywheel effect is surprisingly robust.

[1] https://arxiv.org/abs/2508.03385

I saw that, but the approach taken is questionable (do LLMs represent realistic behavior for scenarios they’ve not been trained for?) and it also doesn’t seem like anything like my suggestions here were tested. It’s better than nothing, but far from conclusive in my opinion.