Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics
This work addresses the challenge of unifying disparate control methods for large language models, which is incremental as it builds on existing techniques to provide a conceptual framework and analysis.
The authors tackled the problem of understanding and comparing different methods for controlling large language models by proposing a unified framework that frames interventions as dynamic weight updates, revealing a consistent trade-off between preference (target concept alignment) and utility (coherent generation). They introduced a new steering approach, SPLIT, which improves preference while better preserving utility, though specific numerical gains are not detailed in the abstract.
Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model's valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md.