When Should an LLM Steering Edit Depend on the Input?

TL;DR

This paper shows that LLM steering edits should depend on input only when the target output varies within a class, and proposes a spectrum of input-dependent edits from fixed direction to query-specific retrieval. It demonstrates that retrieval-based edits outperform fixed-direction steering when within-class variation exists, with optimal input-dependence varying by task granularity and network depth.

Why It Matters

The work provides a principled, measurable criterion for when to use input-dependent steering, improving edit accuracy and efficiency in real-world applications like text correction and content generation. It guides practitioners on how and where to apply edits, reducing reliance on trial-and-error tuning.

Key Findings

• Input-dependent steering is necessary when the desired output varies within a class, measurable via class-level state statistics.
• Retrieval-based edits (using k-nearest neighbors) outperform fixed-direction edits across all layers and tasks when within-class variation exists.
• The required level of input-dependence depends on edit granularity: word-level edits benefit from early-layer retrieval with few neighbors, clause-level edits work well at any depth using class means.