Explain Haiku’s Lack of a Self-Edit/Cross-Edit Difference

Determine why Claude Haiku, unlike the open-weight language models evaluated in the cross-editing study, shows little difference in edit distance between code it generated itself and code generated by other models.

Background

The paper compares self-editing, in which an editor modifies code produced by the same model, with cross-editing, in which the editor modifies code produced by another LLM. The open-weight editors generally make smaller edits to their own implementations than to foreign implementations, whereas Claude Haiku does not exhibit the same pattern: its smallest edits occur on GPT-OSS implementations rather than on its own code. The reason for this discrepancy is left unresolved and identified as a direction for future work.