Skip to content

Comment on We tested 20 LLMs for ideological bias, revealing distinct alignments

Comments

The large differences between gemini-2.5-pro and the gemini-X-flash and gemma models is surprising. It looks like distillation causes an ideological shift. Some, but not all of the other distilled models also show that shift.

Pet theory: distillation causes roughly-random changes, and political alignment wasn't the most important part of the evals for which distill got released, coding skills etc were.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.