New paper builds a scientific theory of atomic features: SAEs of all sizes share features
jessicadai_ · x · 2026-10-07
A new paper introduces a scientific theory of atomic features with mathematically derived, testable predictions. Experiments challenge conventional wisdom: SAEs of different sizes and training data share many features, and large SAEs recover both parent and child features, suggesting hierarchical structure.
More from Research
- 20-year math researcher on the LLM theorem drop: reshuffling vs genuinely new ideas — littmath · 2026-10-07
- Hexagon: a new repository for LLM-assisted math results launches on Terry Tao's blog — alexbilz · 2026-10-07
- OpenAI's Math Breakthrough List Skips Cryptography, Researcher Notes — thomasahle · 2026-10-07
- AI Research Agents Have Radically Different Styles—Leaderboards Reduce Them to One Blind Number — ChengleiSi · 2026-10-07
- Gary Marcus amplifies claim that Lean deserves equal credit with AI for math breakthroughs — GaryMarcus · 2026-10-07
- Diffusion LM paper I-DLM lands COLM 2026 Oral Spotlight, presenting Thursday — james_y_zou · 2026-10-07