Vijay Vankadaru

Vijay Vankadaru

MS Data Science, UC Berkeley
Mechanistic interpretability of LLMs and VLMs
Machine Learning Engineer, DASION

A Recipe for Multimodal Multi-Instance Depression Detection: Frozen Teachers, Schedule-Aware Training, and Audio Contribution Diagnostics

Under reviewACL Rolling Review

V. Vankadaru

Solo-author paper from work at DASION.

Published DAIC-WOZ results are hard to reproduce: preprocessing is under-specified, seeds are missing, repositories are dead, and several reported numbers depend on bias the models should not be exploiting. This paper builds a modern multimodal multi-instance pipeline on top of frozen domain-finetuned encoders, reaches F1 0.87 as an honest bias-aware ceiling, and introduces the audio occlusion delta, a diagnostic that exposes modality collapse that accuracy alone cannot see.

Cite as: Vankadaru, V. A Recipe for Multimodal Multi-Instance Depression Detection: Frozen Teachers, Schedule-Aware Training, and Audio Contribution Diagnostics. Under review.