Safety

Empirical research on content moderation is fundamentally constrained by the opaque deployment of moderation systems on major social …

A benchmark and evaluation framework comparing instruction- vs. example-driven VLM approaches for automated content moderation policy operationalization.

We conduct a comprehensive benchmark study evaluating LLMs for demographic-targeted social bias detection in raw text data, revealing that while certain configurations show promise for scale, significant performance gaps persist across complex social categories.

We explore improving the practicality of providing causal recourse explanations through a novel neural network model-based automation framework.

We explore the helpfulness of unlabeled data for fair, optimal and stable decision-making in societal settings.

Project that was conducted as part of my Master’s thesis to explore the application of generative models to compute counterfactuals for fairness.

This project explores the case for potential bias in generative models such as variational autoencoders. The project also briefly looks at ways to mitigate such bias.