A study introduces a method for steering language models by editing learned B-spline attention surfaces.

What it means

This paper introduces 'attention manifolds', a method to steer language models by editing learned 2D B-spline surfaces that modulate attention values per dimension. The authors show that modifying these smooth surfaces can control what the model attends to, offering a new fine-grained approach to model steering without retraining.

Accuracy & sourcesGrounded

The claim is from the actual arXiv paper (cs.LG, 2610.00257), a primary source. The paper is publicly available and the method is described in detail, so the claim is as solid as the paper itself.

How new

Genuinely new; published on arXiv on 2 October 2026.

Go deeper

• arXiv cs.LG — first to publish