Beyond fixed-turn, pre-scripted evaluations, we find that social bias in multi-turn LLM interactions can emerge late, fluctuate across turns, and re-emerge. These results highlight the need to evaluate social bias as a dynamic phenomenon.
Paper: https://t.co/GISKtZiH9L
[2/2]
1st paper w/@MasahiroKaneko_@dai0NLP@Bollegala@chokkanorg
We study social bias in response-conditioned multi-turn LLM interactions, where follow-up user queries adapt to the evolving dialogue which can vary in length.
[1/2]