Research Scientist · Frontier Model Evals, Alignment Robustness & Trustworthy AI
I build evaluation and intervention methods for knowledge control in ML systems, with a focus on machine unlearning, predictive multiplicity, and inference-time alignment.
My research studies how ML systems store, expose, forget, and use knowledge. I build evaluation and intervention methods that reveal hidden failure modes: residual knowledge after unlearning, decision and explanation inconsistencies across equally valid models, and unreliable inference-time steering with external knowledge, context, or rewards. Across these settings, I aim to make model behavior more auditable, robust, and interpretable, building on my earlier work in privacy, fairness, and information-theoretic ML.
Machine unlearning and residual-knowledge auditing: removing unwanted information from models and testing whether deletion remains robust under perturbations, attacks, and distribution shift.
View research portfolioRashomon-effect research on predictive multiplicity: understanding when equally accurate models, decisions, or circuits disagree, and why this hidden instability matters for fairness, privacy, and interpretability.
View research portfolioInference-time alignment and knowledge control: steering which knowledge a model uses through reward-tail-aware selection, sparse feature interventions, and future memory-unit representations.
View research portfolioShows that models can pass standard forgetting checks while retaining residual knowledge under perturbed deletion queries.
Develops unlearning methods for image-to-image generative models, including diffusion models, VQ-GAN, and masked autoencoders.
Improves the scalability of NTK-based machine unlearning with parameter-efficient updates for large-scale models.
Analyzes and mitigates conflicting individual predictions among similarly accurate gradient boosting models.
Uses dropout-based exploration to estimate predictive multiplicity efficiently without repeatedly retraining neural networks.
Shows that group fairness interventions can improve aggregate fairness while increasing individual-level arbitrariness.
Reveals predictive multiplicity as a hidden cost of privacy-preserving randomized training procedures.
Introduces Rashomon Capacity to quantify score-level disagreement among near-optimal probabilistic classifiers.
Studies predictive multiplicity through neighboring datasets and the sensitivity of model decisions to data perturbations.
Uses reward-tail behavior to adaptively balance optimistic and pessimistic selection in inference-time alignment.
Studies watermarking in low-entropy generation regimes where naïve watermarking can be brittle or distortive.
Detects hybrid human/LLM-written text rather than assuming documents are fully human- or machine-written.
Probes hallucination through perturbation-driven analysis of internal knowledge and model behavior.
Protects private attributes while preserving downstream utility through stochastic data substitution.
Uses information-theoretic transformations to suppress multiple protected attributes while retaining utility.
Develops fairness interventions for multi-class prediction using information projection.
Extends correspondence analysis for representation learning and machine-learning applications.
Regularizes classifier projections to improve continual learning under sequential tasks.
Develops information-theoretic tools for auditing and constraining privacy leakage.
Leading work on adversarial evaluation, unlearning verification, reward and decision instability, and model safety governance.
Developed evaluation methods for residual knowledge, LLM watermarking, partial-LLM text detection, hallucination probing, and predictive multiplicity.
Worked on information-theoretic tools for fairness, privacy, representation learning, and decision uncertainty.
Thesis: Information-Theoretic Tools for Machine Learning Beyond Accuracy — fairness, privacy, and decision uncertainty.
Best Master’s Thesis Award and National Young Scholar Best Paper Award.