Research

Chest studies plotted on the emphysema and heart failure finding words, separating as they move from the input to the output of the language model
Oct 14, 2026Epsilon Health Research Medical AI

How Image Tokens Learn the Language of Radiology

We read the image tokens of our report-drafting model through its own clinical vocabulary and with linear probes. Deeper in the language model, they come to point at the right findings, with no direct supervision on the image tokens.

Aakash Mishra, Arjun Karpur

Report-model attention moving across a hand radiograph as each finding is written
Jul 29, 2026Epsilon Health Research Medical AI

Auditing Our Models for Visual Explainability

A heatmap can reflect what a model computed and still point at the wrong anatomy. We score our models' explanations on localization and faithfulness, and find that learned attention holds up on both.

Aakash Mishra, Arjun Karpur

System diagram: an upstream causal model informs an online adaptive tuning engine that sets ad load under constraints
Dec 2025NeurIPS 2025 Workshop on Constrained Optimization for ML Systems & optimization

Causal-Informed Hybrid Online Adaptive Optimization for Ad Load Personalization in Large-Scale Social Networks

Combines constraint-safe primal-dual control with trust-region Bayesian optimization, guided by a causal model, to personalize how many ads people see. Tested in live A/B tests on a billion-user social network.

Aakash Mishra, Qi Xu, Zhigang Hua, Keyu Nie, Vishwanath Sangale, Vishal Vaingankar, Jizhe Zhang, Ren Mao