Neel Nanda pushes back on LRH critique: one steering vector, four norms, four behaviors
NeelNanda5 · x · 2026-11-08
John Hewitt's new blog post, Do Not Rely on the Linear Representation Hypothesis, shows a single direction in LLM activation space — applied as a steering vector at four different norms — can trigger a different model behavior at each norm.
Neel Nanda responds that this doesn't obviously contradict the Linear Representation Hypothesis: LRH never claims every direction is monosemantic, only that interesting directions decompose into sparse combinations of monosemantic ones — so a+b+c of unrelated monosemantic vectors is likely not monosemantic. He also asks whether the vector was found via optimization.
A substantive methodological debate for anyone doing steering-vector work: distinguish single monosemantic directions from linear combinations.
More from Research
- Latent Sparse Attention paper: indexing eats 98% of prefill at 10M tokens, LSA speeds one layer 22.7x — burny_tech · 2026-11-08
- Double-exponential fit suggests model time horizons could hit ~80 work-years by end of 2027 — scaling01 · 2026-11-08
- ACM policy lets authors share anonymous reviews, no barrier to crowdsourcing — IanArawjo · 2026-11-08
- CHI researcher proposes open peer-review database, wants LLM reviewers banned — IanArawjo · 2026-11-08
- AI capability growth as a tug of war: recursive self-improvement vs diminishing returns — CFGeek · 2026-11-08
- Agent S4 lets agents evolve reusable computer-use interfaces from their own task experience — burkov · 2026-11-08