One steering vector, four norms, four behaviors: don't trust linear representation
johnhewtt · x · 2026-11-05
John Hewitt (Columbia) published 'Do Not Rely on the Linear Representation Hypothesis': a single steering direction applied at four different norms induces four distinct behaviors (sycophancy, French, code style, negative sentiment), breaking the intensity-as-scaling assumption.
Related event: Columbia's Hewitt Warns Against the Linear Representation Hypothesis(3 posts)→
More from Research
- Weight space programming spawns a Rube Goldberg machine where parts have tiny brains — _ScottCondron · 2026-11-05
- Claude read all 3,096 ASHG abstracts to surface 33 human genetics highlights — anshulkundaje · 2026-11-05
- Stanford's James Zou on AI agents entering the wet lab, in Nature — james_y_zou · 2026-11-05
- Hamel Husain: Benchmarks are used like credit scores, so criticism is warranted — HamelHusain · 2026-11-05
- ArkenOS now runs on Rust, freeing room for more parallel agents — _AustinCalvert_ · 2026-11-05
- CoRL paper GUILD: actively guiding robot data collection toward useful auxiliary tasks — RobobertoMM · 2026-11-05