Peer-reviewed publications and preprints. Google Scholar →
A component-based pipeline that audits validity flaws in LLM-agent benchmarks — decomposing tasks into user, environment, ground-truth, and evaluation components. It matches expert judgments at 0.79–0.87 F1 across 6 benchmarks and reshuffles 63% of model rankings on a 30-LLM leaderboard.
A causal method that recovers directional brain-connectivity dependencies invisible to correlation-based graphs, improving downstream diagnostic prediction across ~8K patients via Granger-directed graphs.
A state-based Transformer that models temporal transitions across latent functional-connectivity states, outperforming static-connectivity baselines on clinical prediction.
Preprints and manuscripts under review will appear here.