Elektrine lite

← Feed

@trochee@dair-community.social

Post #1846526

2026-04-19 14:45 UTC

This study's results suggest that LLM sycophancy amplifies publication bias (the "file drawer" problem) The LLMs didn't like to tell the user that their hypothesis was not confirmed by the data H/t @bibliolater https://social.vivaldi.net/@bibliolater/116252085690032777

Replies (1)

  • @KimSJ I'm anthropomorphizing, yes. It's a useful tool to talk about interactions among complex systems I am not attributing actual sentience to the LLMs. I did not say "LLM training data patterns are biased in favor of the appearance of agreement with the user, even when the user has what would be externally deemed a 'bad idea'" instead, I chose a shorthand that uses a concept we already have for describing human-to-human communication "LLMs have a sycophancy problem" @bibliolater

    Open ##2089219