A Control Knob Is Not a Concept: What a Working Steering Vector Licenses, Why Its Equivalence Class Is Set by the Readout That Scores It, and the Controls That Would Upgrade Control to Representation

Abstract

Adding a vector to a language model's residual stream and watching a behaviour change has become the field's most common argument that the model represents a concept along that vector. Refusal directions, persona vectors, truth directions and instruction vectors are all reported this way. A second body of work, mostly from 2025 and 2026, finds that geometrically unrelated vectors produce the same behavioural shift, that refusal is carried by several directions and multi-dimensional cones rather than one, that random directions of activation-scale norm can break safety behaviour, and that the direction which best detects a property is often a poor direction for controlling it. In this analysis the author reviews and synthesises published results to show that steering's defenders and critics have measured different things. A steering result is an estimate indexed by six choices - the contrast set, the extraction rule, the injection site, the magnitude, the readout that scores the behaviour, and the rule that selected the vector from its candidates - and it is asked to carry six claims: control, specificity, necessity, uniqueness, content and use. The paper argues three points. First, the equivalence classes behind the non-identifiability results are set by the readout: the papers that report equivalence under a trait classifier or a binary refusal judge also report, under next-token logits or under the framing of the refusal, differences those readouts hide, so coarse scoring manufactures part of the equivalence. Second, adding a vector and removing it are not symmetric tests: addition is many-to-one, while ablation discriminates between directions, and the single-direction refusal claim was contested on the ablation side. Third, reading and writing come apart in both directions. The paper consolidates the measured values in one table and one figure, states an upgrade procedure from control to representation as an algorithm, and names five studies that would settle the open part. On the record, control is well supported, necessity is supported for a few behaviours, and uniqueness is supported for none.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 83 papers by Pranay Mahendrakar, or his ORCID record.