Columbia's Hewitt Warns Against the Linear Representation Hypothesis

Columbia's John Hewitt argues in a new blog post that interpretability should not rely on the Linear Representation Hypothesis: the same steering vector triggers different behaviors at different norms, and combining individually effective vectors can steer toward unrelated tasks, suggesting activation-space interpretation is fundamentally nonlinear.

2026-11-05 ~ 2026-11-05 · 3 related posts

• One steering vector, four norms, four behaviors: don't trust linear representation — johnhewtt · 2026-11-05
• Hewitt: summed steering vectors can trigger an unrelated third task — johnhewtt · 2026-11-05
• Interp researcher: activation space readings are radically nonlinear, breaking additivity intuitions — johnhewtt · 2026-11-05