A toy example of why RL fine-tuning concentrates mass on max-reward tokens

jessi_cata · x · 2026-09-12

jessicata offers a simple toy example of RL fine-tuning's fixed point: sample a model's next-token distribution 1000 times, score tokens by some f(token) in [0,1], keep samples with probability f(token), then fine-tune the model toward the remainder. Iterating this drives nearly all probability mass onto tokens with maximum f — the reward-maxing fixed point.

Related event: RL Researchers Debate: REINFORCE as Both Policy Gradient and Synthetic Data Method(5 posts)→

Original post →

More from Research

Research channel →