Built independently by an author, for readers. Read the story and support ChapterPal
ChapterPal Logo
Log inSign up

keyword

steer matrix

A steer matrix is a learned transformation matrix applied to internal representations or word embeddings in a language model to guide its generation style, behavior, or task performance without altering the core parameters of the underlying model. By linearly modifying activation or embedding spaces during processing, the steer matrix directs model outputs toward specific attributes, such as designated topics, stylistic preferences, reduced toxicity, or targeted task behaviors. This mechanism offers a parameter-efficient approach to controlled text generation, enabling targeted interventions that can often be scaled, combined, or transferred across models to achieve nuanced multi-attribute control.

1 item

Word Embeddings Are Steers for Language Models

Word Embeddings Are Steers for Language Models

Chi Han, Jialiang Xu, Manling Li, Yi Fung, Chenkai Sun, Nan Jiang, Tarek F. Abdelzaher, Heng Ji

Why you should read this

Proposes LM-Steer, a lightweight approach that applies linear transformations to output word embeddings to achieve parameter-efficient, transferable, and continuous style control across language models without retraining internal weights.

Recent studies have shown that large language models (LLMs), when augmented with additional weights through adapters—for instance, LoRA—can learn specialized tasks without updating model parameters—a paradigm known as either Instruction Tuning (IT) for individual instructions or Task Arithmetic (TA) for task-specific adaptation. In such scenarios, different instruction/task datasets induce distinct steering vectors within LLMs' embedding space, shaping behavior according to downstream objectives. Inspired by recent advances in model merging or parameter composition, we investigate whether multiple steering directions (each optimized toward generating responses on one dataset while discouraging outputs outside its distribution via unlearning-like objectives) can be combined into a single vector enabling zero-shot generalization across diverse tasks.

Added

2026-10-02