Will RL work for LLMs "spill over" to the rest of RL by 2026?
➕
Plus
5
Ṁ409
2026
40%
chance

RL is important for training LLMs and it seems likely that there will be significantly more investment in RL by the major LLM groups this year. Will any of the advances they make be:

  1. Published (any publication that allows the research to be used elsewhere counts, this does not have to be a paper)

  2. A significant advance for the rest of RL

For example, a new version of PPO that is close to SOTA for agents in Atari environments would resolve this YES.

What counts as a "significant advance" is mostly subject to my inscrutable whims, but is aimed more at cool research than important result. Think "very exciting to see at a conference" rather than "revolutionizes the field".

Get Ṁ1,000 play money