Steering Along Manifolds to Control Neural Networks

{
  "doi": null,
  "year": null,
  "title": "Steering Along Manifolds to Control Neural Networks",
  "authors": [],
  "abstract": "Intervening on a model's internal representations to steer behavior, known as representation steering, promises lightweight, adaptable, and granular control of neural networks. While the typical approach uses linear steering by adding a scaled steering vector to hidden representations, this paper argues that steering along manifolds—curved surfaces in representation space—provides better control. Using the cyclic concept geometry of days of the week as a case study, the authors demonstrate that geometry-aware steering reveals a deep connection between the geometry of neural network behavior and representation.",
  "arxiv_id": null,
  "pdf_status": "available",
  "openalex_id": null,
  "source_file": "steering.md",
  "ingested_via": "ingest_one_url",
  "fulltext_path": "/Users/antonborzov/Documents/Research.nosync/papers/steering.md"
}