Determine the property governing the localized-editing regime

Determine what property of an eligible language model determines whether an interpretability-guided localized weight edit will outperform behavior cloning or imitation-based editing, including why the regime boundary varies across model families and can pass between full-pipeline seeds of a single model.

Background

The paper evaluates whether restricting LoRA weight updates to a causally implicated residual-stream band provides an advantage over behavior cloning with updates distributed across all layers. Across five eligible model families, the outcome is heterogeneous: localized editing is a stable success on Qwen2.5-3B, generally fails in different ways on three other families, and crosses the pass/fail boundary across seeds for Phi-4.

The authors test inexpensive predictors based on the cosine between behavior and style directions and on the causal peak's depth, but both prediction rules are falsified by out-of-family tests. Consequently, the paper maps the regime empirically without explaining its determining property and identifies this unresolved explanatory question as its most important open question.

References

The regime boundary of \S\ref{sec:map} is mapped, not explained --- $n = 5$, with the boundary passing through one model's full-pipeline seed variance --- and we deliberately declined to answer ``what property determines the regime'' with a third data-fitted rule; we consider it the paper's most important open question.

— Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures  (2610.21708 - Zavattari et al., 15 Oct 2026) in Section Limitations