Research

Each plate runs in three panels: the problem, the machinery, and the result. The objects carry across the panels, so a single mark can be followed the whole way through. Work done inside a role is written up under that role on Experience.

t0t1t2t3?event in this intervalinterval censored · no continuous monitoringbfirst passagelearned drift · Arrhenius clock · Fokker–Planck transitiontimehistorytoday80% credible intervalP(event by day h | history)days ahead
The subject is observed at four irregular visits. The event happens after the last one, at an unknown time.Each observation pins a latent diffusion whose clock is set by the conditions. The event is its first passage across the boundary.Solving the Kolmogorov equations on a grid gives the event-time density exactly, so the reported interval is calibrated rather than nominal.

Exact First-Passage Likelihoods for Neural Diffusions under Interval Censoring

University of Maryland · first author, in preparation

Forecasts time to an irreversible event from observations taken at irregular visits, which is how food spoils, equipment degrades and disease declares itself. The latent state is a neural diffusion whose clock is a physical function of the conditions; the event is its first passage across a learned boundary, and the event-time law is solved on a grid from the Kolmogorov equations rather than sampled. The likelihood is proper under interval censoring, so the reported interval is calibrated rather than nominal.

neural SDE · first passage · Fokker–Planck · interval censoring · JAX · Diffrax

Repository
easymediumhardone label, very different problemsP(solved)easyhardsix datasets · mathematics, programming, chess, reasoningmodel Amodel BB falls off hereaccuracydifficulty
Benchmarks record difficulty in three buckets. Problems inside one bucket are not remotely alike.Item-response theory and Glicko-2 re-estimate every problem on one continuous standardized scale, from human and model attempt records.Accuracy against that scale is a curve, not a scalar. It shows exactly where each model stops generalizing.

Standardized Difficulty Labels for Profiling LLMs

UMIACS · NeurIPS 2024, Datasets & Benchmarks

Replaces coarse difficulty buckets with a continuous, standardized difficulty score for every problem across six datasets, estimated by item-response theory and Glicko-2 from human attempt records and model evaluations. Accuracy against that scale is a curve rather than a scalar, which is what makes easy-to-hard generalization measurable at all.

item-response theory · Glicko-2 · difficulty calibration · LLM evaluation

Paper Repository Dataset
greenhazardteea two-player turn-based stochastic gameopponent-blindconditionedhazardvalue iteration over match state × lie × distanceone narrow bandno gaingainmatch state · holes remaining × margin
Match play is scored by hole, not by stroke. A blow-up costs one hole, and a birdie wins exactly one.Two policies leave the same tee: one solved blind to the opponent, one solved against their shot distribution.Across every match state the gain is near zero. It exists in one narrow band, and opponent statistics that precise are not obtainable.

Should Professionals Consider Their Adversary's Strategy?

Computational Statistics, 2024 · first author

Match play in golf, posed as a two-player turn-based stochastic game and solved as a Markov decision process on PGA Tour ShotLink data. Conditioning on the opponent is advantageous in a narrow set of match states and is otherwise dominated by optimal stroke-play strategy, since opponent statistics precise enough to act on are not obtainable in practice. The method extends to any two-player sport scored in turns.

Markov decision processes · stochastic games · value iteration · ShotLink

Paper Preprint Repository
at triageadmitteddischargedinsuranceESIethnicityagegenderlanguage0insuranceESIethnicityagegenderlanguage← dischargedadmitted →matched on Emergency Severity Indexprivate planfederal plana gap severitydoes not explainadmission rate
Every arrival is sorted at the door: admitted, or sent home. Six attributes are on the record when that happens.Shapley values weigh what each attribute contributed. Insurance status carried more than the severity index itself.Matched on severity, admission rate still differs by insurance program. Language, against expectation, was not a barrier.

Using Pattern Recognition to Detect Implicit Bias in Hospital Triaging

Annals of Biomedical Science and Engineering, 2023 · co-author

Classifies hospital triage dispositions from the attributes recorded at intake, then attributes each decision with Shapley values. At a given Emergency Severity Index, patients on federally run insurance programs were admitted less often than patients on private plans, and ethnicity weighed more heavily on patients under forty. Language was not a barrier.

Shapley attribution · balanced random forest · clinical records

Paper Repository
ScholarLinkedInGitHub
© Nishad Wajge 2026