It’s a common misconception, I think, that LLMs don’t remember what happened / what they saw in RL training. The gradient update is with respect to the model’s full inner states!!! I think the Opus…

Posters: repligate. repligate

2 likes recorded by the source corpus.