Research | Jaden Moon

Research

My work asks when a model should be trusted, when it should defer, and what evidence supports that decision. The projects below span wearable sensing, multimodal learning, medical imaging, and rigorous failure analysis.

Theme 01

Selective prediction

Methods that decide whether a prediction is sufficiently reliable to surface under participant variation, dataset shift, and asymmetric clinical risk.

Accepted · WellComp 2026 01

ReliaGate

When should a wearable stress model surface a prediction—and when should it withhold one?

+10.1 pp coverage −5.6 pp accepted risk 4 public datasets

Ongoing · Mayo Clinic 03

Radiotherapy Segmentation QA

How can an AI system identify autosegmentations that require human review under asymmetric clinical risk?

59.4% low-risk coverage 99.91% NPV 0.936 AUROC

Theme 02

Reliability diagnostics

Interventions and follow-on methods that test whether reliability signals actually influence decisions—and how that dependence changes over time.

Accepted · INTERSPEECH 2026 02

Multimodal Reliance Diagnostics

Does a quality-aware fusion model actually use its reliability signals at inference time?

25 runs / dataset 2 multimodal datasets

Ongoing follow-on work 04

ReliaGate++

Extending static routing into validation-locked temporal emission control and episode-level reliability auditing.

Cue precision Subject reach Episode coverage

Earlier work

Methods and systems that shaped the agenda

Earlier projects in language-model adaptation and ambiguous medical-image segmentation that shaped my focus on evaluation and failure modes.

Prior research · AIM HIGH Lab 05

Therapeutic Dialogue Adaptation

How can open language models be adapted and evaluated for therapeutic dialogue under constrained compute?

1,753 responses evaluated 4-bit quantization

Course research · Dartmouth 06

LDASeg

How should a segmentation model represent multiple plausible annotations?

Best class project citation