One commenter asks for more than anecdotes; another replies that earlier material shows reinforcement learning can teach facts.
One commenter says they would like more than anecdotes. In a reply, another asks what further evidence could be offered and argues that the material discussed earlier explicitly shows reinforcement learning can teach facts.
394
6 Sources, first seen ago
@LChoshen@yoavgo What more can you get? The above explicitly shows you can rl to learn facts
@yoavgoi think that "being able to learn facts via RL (or SFT)" was never under dispute. of course it is possible to learn facts this way. but in "early" LLMs it was quite clear that these stages had a different role, and learning facts there was considered "a failure" of sorts. this is no longer the case. and i am curious to learn what is the new balance. how much new knowledge is added in post training these days? which kinds of knowledge? which stages of post training? etc@LChoshen@yoavgo What more can you get? The above explicitly shows you can rl to learn facts@yoavgoi think that "being able to learn facts via RL (or SFT)" was never under dispute. of course it is possible to learn facts this way. but in "early" LLMs it was quite clear that these stages had a different role, and learning facts there was considered "a failure" of sorts. this is no longer the case. and i am curious to learn what is the new balance. how much new knowledge is added in post training these days? which kinds of knowledge? which stages of post training? etc@LChoshen@yoavgo What more can you get? The above explicitly shows you can rl to learn facts@yoavgoi think that "being able to learn facts via RL (or SFT)" was never under dispute. of course it is possible to learn facts this way. but in "early" LLMs it was quite clear that these stages had a different role, and learning facts there was considered "a failure" of sorts. this is no longer the case. and i am curious to learn what is the new balance. how much new knowledge is added in post training these days? which kinds of knowledge? which stages of post training? etc@LChoshen@yoavgo What more can you get? The above explicitly shows you can rl to learn facts@yoavgoi think that "being able to learn facts via RL (or SFT)" was never under dispute. of course it is possible to learn facts this way. but in "early" LLMs it was quite clear that these stages had a different role, and learning facts there was considered "a failure" of sorts. this is no longer the case. and i am curious to learn what is the new balance. how much new knowledge is added in post training these days? which kinds of knowledge? which stages of post training? etcβ
Not ranked yet
β
Not ranked yet
Comments