DeepMind's Yasaman Bahri to lecture at Harvard on how language statistics shape LLM representational geometry
yasamanbb · x · 2026-11-04
Google DeepMind researcher Yasaman Bahri speaks at Harvard's CMSA 'New Technologies in Mathematics' seminar (Nov 4, 2pm ET, via Zoom) on the links between representational geometry and natural language statistics.
- LLMs organize internal representations into rich geometry: categorical concepts map to linear directions, while words with spatial/temporal semantics trace winding manifolds
- These geometries emerge across diverse architectures and training algorithms, and are key to understanding model computations
- The work builds a theoretical framework from classic word-embedding methods and co-occurrence statistics of natural language, with predictions matching empirical findings
- It offers a unified explanation for prior observations of representational geometry and a rigorous foundation for further theory
More from Research
- ORNL Paper: Useful Quantum Computing Needs 4M Qubits and 800 Square Meters — anirbanbandyo · 2026-11-04
- DR Pixel: open-source 27B deep research agent reads paper pages without OCR, matches GPT-5.4 — CShorten30 · 2026-11-04
- Nigerian Legal Corpus released: cleaned constitutional and legal texts for RAG and legal NLP — saheedniyi_02 · 2026-11-04
- Vulnerability research is a topology problem: reusing local proofs across audits — dyn___ · 2026-11-04
- Agents will soon edit their own weights and memory — but who studies when to use which? — sarahcat21 · 2026-11-04
- Hand-writing all weights of a 100k-param model beats training it at spotting fake legal citations — AaronBergman18 · 2026-11-04