An AI-MI–supported study proposes a new architecture for language models that externalizes factual knowledge to a knowledge base rather than storing it in model weights. “Co-LMLM: Continuous-Query Limited Memory Language Models” by Yair Feldman, Linxi Zhao, Nathan...
Publications
Separating Prediction and Memory Improves Transformers
An AI-MI–supported study proposes that standard Transformers overload a single computation stream to both predict the next token and store state for future tokens—and that disentangling these roles yields more efficient, higher-performing models. Giovanni Monea,...
AI-MI Research Identifies Anisotropy Key to Magnon-Mediated Superconductivity in 2D Ferromagnets
An AI-MI–supported theoretical study shows that a small easy-plane magnetic anisotropy is the key ingredient enabling magnon-mediated superconductivity in 2D itinerant ferromagnets. Vladimir Calvera, Heqiu Li, Yijie Wang, B. Andrei Bernevig (Princeton), and Andrey V....
AI-MI Researchers Release a Graphlet-Histogram Database of 149,082 Inorganic Crystals
AI-MI researchers have introduced Graphlet-MP, a database that encodes inorganic crystal structures as graphlet histograms — a data-efficient, interpretable, machine-learning-ready representation aimed at accelerating materials discovery. Covering 149,082 inorganic...
AI-MI Follow-Up Discovers New Superconductors in the PtPb₃Bi Structure Type
A follow-up from AI-MI’s superconductivity effort reports a new family of superconductors sharing the PtPb₃Bi structure type — extending the AI-prediction-driven discovery that produced PtPb₃Bi itself, featured as the June AI-MI research highlight. The team...
AI-MI study tests large language models as “world models” for high-Tc superconductivity
The first study out of the AI Materials Institute, led by Director Eun-Ah Kim, tests how well large language models grasp the scientific literature.