Study 1b: This One Weird Trick does NOT cause incorrectness cascades
aizi.substack.com
[Edit: The data collected for this study was produced by critically bugged code. Please see bug writeup here and the results here. Please consider this study as retracted.] [This post is based on my preregistration here.] Abstract Following up on previous work, I found that the tendency of GPT-3.5 to strongly prefer factual answers is not significantly affected by changing answers from “multiple choice” to “true/false”.
Study 1b: This One Weird Trick does NOT cause incorrectness cascades
Study 1b: This One Weird Trick does NOT cause…
Study 1b: This One Weird Trick does NOT cause incorrectness cascades
[Edit: The data collected for this study was produced by critically bugged code. Please see bug writeup here and the results here. Please consider this study as retracted.] [This post is based on my preregistration here.] Abstract Following up on previous work, I found that the tendency of GPT-3.5 to strongly prefer factual answers is not significantly affected by changing answers from “multiple choice” to “true/false”.