The RLM cult is now shilling an obviously good idea
Train LLMs to decompose problems into code vs in-weights
This is similar to RL environments, except the focus is on decomposing problems rather than solving problems
www.alphaxiv.org/abs/2611.har...
1 Boost
0 Likes
the idea is basically, fine tune a laptop-sized LLM to be exceedingly good at operating a badass harness, and you won’t be sad