VISION is a framework for robust and interpretable code vulnerability detection using counterfactual data augmentation. It leverages GNNs, LLM-generated counterfactuals, and graph-based explainability to mitigate spurious correlations and improve generalization on real-world vulnerabilities (CWE-20).

A Framework for Robust, Self-Recovering Tool-Using Language Model Agents — trained on 50K+ failure-annotated trajectories for fault-tolerant reasoning and recovery.

Investigating the "Gradient Noise Paradox" in AI Safety: A study on the conflict between Differential Privacy (DP-SGD) and Adversarial Training. Uses a custom "Shadow Model" pipeline to synchronize Opacus with PGD attacks, demonstrating how privacy-preserving noise systematically degrades model robustness