Activation Steering Vectors
Methods for decomposing, composing, and applying activation-level steering directions to control model behavior.
4 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Nov 2026 · Keio University, RIKEN
AI steering vectors contain distinct functional parts good for control
Decomposes task vectors into recognition, learning, and residual components and shows they can be separately optimized for control.
Nov 2026 · Independent researcher
Above 7B parameters, caving in to pushback can be partly switched off
Identifies a linear activation direction controlling capitulation (abandoning correct answers under pushback) and shows it can be suppressed in large models.
Nov 2026 · Soongsil University
Random tweaks suppress leaked personal data as well as targeted fixes
Tests whether sparse-autoencoder interventions for suppressing PII are truly targeted by comparing with random directions.
Nov 2026 · MATS, Arb Research
Language models can outdo text searches at revealing internal signals
Trains a model to generate text that strongly activates a specified internal signal in another model, outperforming text search.