Large language models — architectures, training, evaluation, and the tools built on them.
Ask a chatbot to fill in a personality test and it will happily oblige, agreeing that yes, it is the life of the party, or no, it rarely feels anxious. The easy reaction is to wave this away as a…
Measuring and Shaping Personality in Language Models Read More »
More and more people are bringing their troubles to a chatbot instead of a friend. Many come away feeling genuinely heard. Not everyone does, and the stakes of getting it wrong in an emotional conversation run high, from unhelpful advice…
Can LLMs Reliably Judge Empathic Communication Read More »
A materials scientist sits down with a fresh stack of papers and already knows the stack is a rounding error. The field now publishes more abstracts in a year than any one person could read in a decade of weekends.…
How LLMs and Concept Graphs Predict New Materials Research Read More »
A synchrotron beamline is a strange place to spend a Tuesday at three in the morning. Beam time is precious, booked months ahead, and the clock does not stop for a tired researcher nudging motors a fraction of a degree…
An Agentic AI That Operates a Synchrotron Beamline Read More »
Ask a chatbot like GPT-4 whether a particular crystal is magnetic and it will give you a confident answer that is often wrong, because it has only ever read about materials, never looked at one. The information that actually decides…
MatterChat Brings Multimodal LLMs to Materials Science Read More »
A researcher at RPTU Kaiserslautern spends a week rewriting the same instruction to a language model forty different ways, tracking which phrasing nudges accuracy up by two points and which one tanks it by ten. That is not a hypothetical.…
Prompt Optimization in NLP Has a Benchmarking Problem Read More »
For years the only way to get a language model to react to an image was to build a small pipeline in front of it. Run OCR to pull out any text, run a separate captioning model to describe the…
Native Vision In Large Language Models, What It Buys You Read More »
A support engineer pastes a customer question into a chatbot built on GPT or Claude, gets back a clean, confident sounding paragraph, and has no real way to know whether that paragraph is grounded in fact or quietly invented. That…
UQLM Scores How Much You Should Trust an LLM Answer Read More »
A model that can read a million tokens sounds like it solves the whole document problem in one move. Paste the entire codebase in, paste the entire contract in, ask your question, done. In practice the win is real for…
Long Context Windows And When They Actually Help Read More »
Ask a text to motion model to simulate someone painting a wall, and it will produce something, a figure with an arm that moves and a brush that gestures. Whether that figure moves the way an actual person painting a…
Why GPT-4 Rewritten Prompts Only Sometimes Improve HUMAN Motion Simulation Read More »