3 ms·
The Reversal Curse: LLMs trained on “A is B” fail to learn “B is A” [pdf]
- mellosouls 3y agoTwitter thread: https://twitter.com/OwainEvans_UK/status/1705285631520407821 https://twitter.com/OwainEvans_UK/status/1705285631520407821 https://nitter.net/OwainEvans_UK/status/1705285631520407821 https://nitter.net/OwainEvans_UK/status/1705285631520407821 Does a language model trained on “A is B” generalize to “B is A”? E.g. When trained only on “George Washington was the first US president”, can models automatically answer “Who was the first US president?” Our new paper shows they cannot!
- gus_massa 3y agoRelated discussion about a blog post about this article (posted a few minutes later) https://news.ycombinator.com/item?id=37617863 https://news.ycombinator.com/item?id=37617863 (34 points | 2 hours ago | 57 comments)
- mellosouls 3y agoTFW the second thread gets a link from the original instead of vice versa and even the effort comment from the OP to the twitter thread and the nitter alternative doesn't get upvoted... /Harumph
- gus_massa 3y agoIt's hard to know which thread would get traction. I noticed this pot a few hours after it was posed and for some weird reason the other thread had many comments that may be relevant. My guess is that the paper makes a technical claim and the blog post makes a big generalization, and the generalization is too controversial and gets a lot of comments.
- mellosouls 3y agoHa yeah no worries cheers, I was half-amused - I've been on the positive side of the Sword of Arbitrary Credit enough times that I can't be too grumpy about it.