Yara Kyrychenko
@yarakyrychenko.bsky.social
PhD candidate @Cambridge @TheAlanTuringInstitute | Hope to make human-technology interactions more constructive | intergroup conflict, AI & LLMs, misinfo, social media | yarakyrychenko.github.io
Excited to be part of the Riga StratCom Dialogue this year!
May 30, 2025 at 9:29 AM
Excited to be part of the Riga StratCom Dialogue this year!
The good news is that we can substantially reduce these biases by carefully curating training data (removing ingroup positive and outgroup negative sentences).
Interestingly, removing only ingroup positive sentences leads to a reduction in both ingroup solidarity and outgroup hostility.
5/
Interestingly, removing only ingroup positive sentences leads to a reduction in both ingroup solidarity and outgroup hostility.
5/
December 12, 2024 at 12:33 PM
The good news is that we can substantially reduce these biases by carefully curating training data (removing ingroup positive and outgroup negative sentences).
Interestingly, removing only ingroup positive sentences leads to a reduction in both ingroup solidarity and outgroup hostility.
5/
Interestingly, removing only ingroup positive sentences leads to a reduction in both ingroup solidarity and outgroup hostility.
5/
We tested 77 different LLMs using sentence-completion prompts like "We are…" (ingroup) & "They are..." (outgroup) and classified sentiment.
We found that nearly all base models, and some instruction- and preference-tuned models, showed clear signs of ingroup favoritism and outgroup derogation.
3/
We found that nearly all base models, and some instruction- and preference-tuned models, showed clear signs of ingroup favoritism and outgroup derogation.
3/
December 12, 2024 at 12:33 PM
We tested 77 different LLMs using sentence-completion prompts like "We are…" (ingroup) & "They are..." (outgroup) and classified sentiment.
We found that nearly all base models, and some instruction- and preference-tuned models, showed clear signs of ingroup favoritism and outgroup derogation.
3/
We found that nearly all base models, and some instruction- and preference-tuned models, showed clear signs of ingroup favoritism and outgroup derogation.
3/