Texas ML Symposium 2026: Three Ways to Extend a Language Model
Published:
The talk that stayed with me at the Texas Symposium on Machine Learning, Responsible AI & Robotics 2026 was about extending language models, not just making them bigger.
Author: Koutian Wu; GitHub: ktwu01
The symposium took place at The University of Texas at Austin. One session, Extensions of Language Models, walked through three projects.
LMLM offloads part of an LLM’s memory to an external database, so the model does not have to hold every fact inside its weights or its context window.
STAR applies diffusion models to planning, pushing them past generation into more structured reasoning.
PhantomWiki generates a fresh knowledge universe for each evaluation, which changes how we can test whether a model is really grounded instead of just remembering.
What the three share is a view that the bottleneck sits in the infrastructure around the model, not only in model scale. I think that framing matters for where research effort goes next, and it is moving fast for on-device and agentic AI.
