Fewer, Better Self-Edits

Self-Adapting Language Models (SEAL) write their own training data (“self-edits”) and apply it to their own weights via LoRA, but leave open how many edits to consume and with which loss. We study both choices across Qwen2.5-1.5B/3B/7B, ending on MIT’s own continual stack under a GPT-4.1 judge. First, at matched optimizer budget a single well-chosen self-edit equals SEAL’s union-of-five (0.395 vs 0.395 on 3B) using 20% of the edit data, and we find no cross-passage interference up to 50 documents in one adapter. Second, replacing SFT with forward-KL distillation toward a frozen base-model teacher that reads the source document cuts prior-knowledge destruction roughly 4× at equal injection on 7B, beating both SFT and replay (final accuracy 0.565 vs 0.464 / 0.457), replicating SDFT’s central finding inside SEAL’s self-edit loop. Third, teacher design does not transfer across scale: the best teacher context at 1.5B is the worst at 7B, and the SDFT-faithful moving self-teacher collapsed outright in our port (preliminary, n=1) while a frozen-anchor variant stayed stable. We further contribute a per-question measurement decomposition, a watch-set protocol with a preregistered validity criterion, a gated reproduction patch for MIT’s stack, and a catalog of negative results and pitfalls; every claim is labeled with the strength of evidence behind it.