AILessWrongabout 2 hours ago
LLM capabilities stem mostly from imitative learning
21 min read
LLM capabilities come mostly from imitative learning, not reinforcement learning from verifiable rewards (RLVR), despite RLVR's prominence. RLVR may use significant training compute but imparts minimal information. Evidence shows non-RLVR models can match RLVR results via sampling or ensembling, and chain-of-thought remains legible, indicating imitative learning's dominance.
Level
Hype check
Tap to vote and see what everyone thinks.
