Research
We study the behavior of language models and the conditions under which their outputs can be evaluated, verified, and used responsibly.
Our research focuses on reasoning, faithfulness, self-correction, calibration, retrieval, and evaluation in practical settings. We publish methods, results, limitations, and negative findings. Each publication includes sufficient information about its data and procedures to support scrutiny and, where possible, replication.
Research publications
Measuring faithfulness in chain-of-thought reasoning
A method for testing whether a model's stated reasoning reflects the computation behind its answer, applied to four open models.
Self-correction helps only when the model can verify
A controlled study of when asking a model to review its own work improves accuracy, and when it does not.
What production deployments taught us about evaluation
Notes from deploying question-answering systems inside real organizations, including the tests we now require before handover.
Prompting strategies that did not improve calibration
A negative result: five commonly recommended prompting techniques produced no reliable improvement in the relationship between model confidence and accuracy.