Simplex Finds 'Telescoping Cone' Belief Geometry in LLM Activations

Simplex (Asterra Institute) published work showing that LLMs trained on mixed-source data exhibit a 'telescoping cone' belief geometry in activations, derived theoretically and confirmed empirically.

2026-09-10 ~ 2026-09-11 · 3 related posts

• Simplex recovers telescoping-cone belief geometry from LLM activations — Hidenori8Tanaka · 2026-09-10
• Simplex research derives transformers' "multidimensional features" activations from theory — livgorton · 2026-09-10
• Simplex publishes LessWrong work showing LLM belief geometry forms telescoping cones — fatihdin4en · 2026-09-11