GoKawiilResearcher derives LLM steering vectors from Qwen3-1.7B's Jacobian lens using only concept tokens
An independent experiment explored inverting the 'J space' Jacobian lens, normally used to decode what an LLM is about to say, to instead generate activation steering vectors directly from a handful of concept tokens. Using Qwen3-1.7B and Neuronpedia's published J lens, the method successfully steered simple behaviors like all-caps output or unusual speech patterns, benchmarked against existing steering vectors and an abliterated refusal-removed model.
GoKawiil's interpretation of the reporting above, not reported fact.
If concept tokens alone can generate usable steering vectors, it could remove the need for expensive finetuned model organisms to control LLM behavior, making activation steering far more accessible. The catch is that this shortcut only holds for simple, tokenizable behaviors and becomes unreliable and hallucination-prone for nuanced or complex traits that resist clean verbal description.
- Jacobian lens can be inverted to build steering vectors from just a few concept tokens, tested on Qwen3-1.7B.
- The technique works well for simple, easily verbalized behaviors like all-caps text but breaks down on complex traits.
- Findings were benchmarked against existing steering vectors and an abliterated model, with code released publicly for replication.
Apple MacBook Pro (M4, 16GB) — The author ran all their Jacobian-lens steering experiments locally on a MacBook, and unified memory is exactly what makes running small LLMs like Qwen3-1.7B on-device practical. A MacBook Pro with an M4 chip gives you a quiet, portable interpretability lab that can load models without a GPU rig.
See Apple MacBook Pro (M4, 16GB) on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: darshanmakwana412.github.io — Darshan Makwana, 2026-09-06
Published there as: “Extracting Steering Vectors from J space”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.