Properties of Privileged-Context Generators Behind Source-Associated Drift

Identify which properties of privileged-context generators account for the source-associated differences in policy drift and parameter-update behavior observed in on-policy self-distillation, using controlled comparisons that separate authorship, model capacity, prompting, and verifier-based selection.

Background

The paper evaluates nine combinations of privileged-context content and source in on-policy self-distillation using Qwen2.5-7B across ProofWriter, MuSiQue, and Big-Math. It finds that source choices are generally associated with smaller differences in drift than content choices, but source-dependent differences remain: self-generated conditions tend to have more closely aligned parameter updates with one another than with the external-source condition.

The observed source effects cannot be attributed to a single factor because the external source differs simultaneously in authorship, model identity, model capacity, prompting, temperature, and selection procedure. The paper therefore leaves unresolved which generator properties are responsible and calls for more granular controlled comparisons to disentangle these contributions.