Azwar Abdulsalam

RL Post-Training Builds Compositional Reasoning Strategies

Azwar Abdulsalam, Nishil Patel, Andrew Saxe

Does RL post-training merely amplify primitive skills already latent in a base model, or can it compose primitive skills into new higher-level strategies? We study this question in a fully observable rewrite-grammar environment where the pretraining distribution is known and ever…Read more