AI2-Israel's Yoav Goldberg argues LLM researchers are rebranding established ML concepts with unnecessary reinforcement learning jargon

@srush_nlp inventing new terminology altogther is also fine, btw. but the case here seem to be just borrowing terminology from RL without getting anything new from it.

I am a huge RL terminology skeptic. Having to teach MDP and POMDP made me sick. And the reward/advantage/q function is a mess. But I think in other ways this is the good case of a field working in isolation for a decade and then having the right abstractions ready to go when LLMs needed them.

@yoavgo @awasthi_a_ rizz Yoav… we do it fir the rizz

@srush_nlp (also i do think it is important to distinguish "policies" where the actions are tokens, from policies where the actions are, say, tool calls or reasoning steps. so i do not think we should abstract over these things..)