Behavior of the KL–forgetting link at frontier scale and across domains

Characterize the behavior of the empirical relationship between forward KL divergence on the new task and catastrophic forgetting at frontier-scale models and in diverse generative domains, and determine whether this relationship persists or changes in these regimes.

Background

The paper demonstrates the KL–forgetting relationship across moderate-scale LLMs and a toy setting, and shows RL’s on-policy bias toward KL-minimal solutions. However, the generality of this relationship at much larger (frontier) scales and in broader generative domains has not been established. The authors explicitly note this gap as an unresolved question.