Steering vector intuitions: persona, concept, belief and task vectors are all the same thing
voooooogel · x · 2026-10-02
In a technical thread responding to the claim that "steering vectors are behavioral biases," the author shares useful intuitions: persona, concept, belief and task vectors are all the same construct, just extracted (mean differences, PCA, NLA reconstruction, SAE) and applied (activation addition, patching) differently; the LLM residual stream is a noisy bag of representations, and steering-vector extraction is about distilling it into clean, reusable, interpretable directions; and mean-difference works by cancelling distractor features on either side of the subtraction — e.g. ignoring "user age" features when extracting a model-distress vector.
More from Research
- New video traces how adversarial objectives evolved beyond GANs and self-play — manicman1999 · 2026-10-02
- Peer-reviewed paper questions whether gene expression model benchmarks are informative — simocristea · 2026-10-02
- Oxford VGG Unveils SynCity 3000, Generating Globally Coherent Scene-Scale 3D Worlds — rsasaki0109 · 2026-10-02
- Christian Szegedy revisits 2019 interview: his 'crazy' timelines for AI math and code — ChrSzegedy · 2026-10-02
- Omni-Embed-Mini: A 0.9B Embedder Adds Five Modalities Without Touching Text Weights — _reachsumit · 2026-10-02
- Apple paper: structured selection-based reasoning cuts search agent latency by 90% — _reachsumit · 2026-10-02