Manifold Steering

Steering Along Manifolds to Control Neural Networks

Concept geometry provides a blueprint for controlling the behavior of neural networks—if you know how to look.

https://www.goodfire.ai/research/manifold-steering

Shared Geometry of Representation and Behavior

Steering along fitted activation manifolds produces smoother behavioral transitions than linear interpolation, with 2.8× lower cumulative Bhattacharyya energy on average across the tested concept tasks. Optimizing activation paths to follow the behavior manifold also recovers the activation manifold’s curvature, supporting a bidirectional geometric relationship across language and visual world models.

arxiv.org

https://arxiv.org/abs/2605.05115

The Confidence Manifold: Geometric Structure of Correctness...

When a language model asserts that "the capital of Australia is Sydney," does it know this is wrong? We characterize the geometry of correctness representations across 9 models from 5 architecture...

https://arxiv.org/abs/2602.08159

Manifold Steering

Shared Geometry of Representation and Behavior

Recommendations