Lav R. Varshney

The paper introduces TraceGuard, a detectability-aware antidistillation method that identifies and poisons 'thought anchors'—sparsely critical sentences—to degrade student model learning without making the defense obvious.

The paper introduces containment verification, a novel method that provides safety guarantees by formally verifying the agentic framework itself, ensuring safety regardless of the underlying AI model's capabilities.

The paper introduces an optimal black-box auditing framework using Donsker-Varadhan estimators to estimate Rényi differential privacy (RDP) guarantees for machine learning algorithms.

This paper introduces a controlled, two-player extension of the Alternate Uses Test (AUT) for comparing human-human and human-AI co-creation under matched conditions, demonstrating equivalent originality with a GPT-4 partner and human partner, and identifying factors influencing performance.