New paper on cost-aware LLM diagnosis at DAIH (COLM 2026)

Adaptive Bayesian Active Querying with LLMs for Efficient Information Gathering was accepted at DAIH: Deploying AI in Healthcare, a workshop at the Conference on Language Modeling (COLM) 2026.

The paper asks a question next door to our disease progression work: rather than reconstructing a biomarker trajectory after the fact, how should a system decide what to ask next? LLMs can generate plausible diagnostic questions and guess how likely each answer is, but they don’t track calibrated beliefs on their own. We pair an LLM with a Bayesian decision layer that scores each candidate question (or costly test) by expected information gain per unit cost, so cheap symptom questions narrow the hypothesis space before expensive labs get ordered. On real emergency department cases, the resulting system reached 86.6% diagnostic accuracy at 31% lower cost than the best non-adaptive baseline.

The project was led by Ognjen (“Ogi”) Malkoc and Shubham Saha of Gaudiy Inc., together with Mizuki Oka of JPCCA and Chiba Institute of Technology, Hongtao Hao of the University of Wisconsin–Madison, Grisha Szep, and our own Joseph Austerweil. Thanks also to JPCCA for their support of this work, as they have for our disease progression modeling.

Code and data are available on GitHub.