Validation at Larger Scales and with Non-LoRA Edit Writers

Validate LadderEdit at scales larger than the evaluated 7B–32B backbones and with edit writers other than LoRA-based writers to determine whether its residual-ladder storage benefits generalize beyond the current experimental setting.

Background

The main experiments evaluate LLaMA-3-8B, Mistral-7B, and Qwen2.5-7B, with supplementary results on Qwen2.5-14B and Qwen2.5-32B. The long-horizon experiments reach 50,000 sequential edits, and the method assumes that each edit is initially acquired as an exact LoRA-style update.

The authors explicitly identify broader validation as unresolved, particularly for substantially larger models and for edit-writing mechanisms that do not produce LoRA updates. Such validation would test whether the observed rank heterogeneity, audit-based promotion, and persistent-memory reductions are properties of LadderEdit itself or of the LoRA-based acquisition setting.