Simplex research derives transformers' "multidimensional features" activations from theory
livgorton · x · 2026-09-10
New work from Simplex asks what structure LLM activations should have, given training data drawn from many sources. Under natural assumptions about data structure, the belief geometry forms "telescoping cones," and transformers are shown to represent this structure. The post offers a theoretical derivation of the interpretability field's picture of activations as sparse linear combinations of multidimensional features, rather than relying purely on empirical observation.
Related event: Simplex Finds 'Telescoping Cone' Belief Geometry in LLM Activations(3 posts)→
More from Research
- Santa Fe Institute opens 2027 Complexity Postdoctoral Fellowships, deadline Sep 30, 2026 — yoavartzi · 2026-09-11
- Are Lean soundness bugs a real risk? Weighing how AI proofs translate to ZFC — jessi_cata · 2026-09-11
- Lean AI proofs to ZFC: researcher says translation is feasible, with caveats — jessi_cata · 2026-09-11
- Novel Reasoning Effort Control Scheme Analyzed: Graded GRPO Training — stochasticchasm · 2026-09-11
- Building a full LLM from scratch over weekends with Stanford's CS336 — stanfordnlp · 2026-09-11
- Universal YOCO paper combines recursive compute with efficient attention for depth scaling — donglixp · 2026-09-11