Experience
Research and professional experience.
Postdoctoral Research Associate
Feb 2025 – PresentLehigh University, Department of Mathematics
- Building an attention-based foundation model that maps observational data to causal graphs, enabling efficient zero-shot causal graph inference (paper under review).
- Developed the Rational Sparse Autoencoder (RSAE), replacing fixed SAE encoder activations with a trainable rational function to improve reconstruction and downstream fidelity for LLM interpretability (ICML 2026 MechInterp Workshop).
- Developed Poly4mer, a physics-augmented chemical language model for polymer property prediction and targeted discovery (NeurIPS 2025 ML4PS Workshop).
Foundation ModelsCausal DiscoveryMechanistic InterpretabilityPolymer Design
Research Scientist Intern
May 2024 – Aug 2024IBM–Rensselaer Future of Computing Research Collaboration
- Designed CircuitLasso, a scalable circuit learning method for mechanistic interpretability of LLMs. Used sparse linear regression to efficiently uncover relationships among high-dimensional sparse-autoencoder features, revealing how interpretable semantic features propagate through the model and improving downstream performance (ICML 2026 MechInterp Workshop).
Mechanistic InterpretabilityLLMsSparse Autoencoders
Research Scientist Intern
Summers 2020, 2021, 2022IBM–Rensselaer AI Horizons Research Collaboration
- Proposed an uncertainty-guided Bayesian inference framework with multi-level uncertainty quantification to improve domain generalization in computer vision (CVPR 2026).
- Developed invariant causal Markov blanket feature learning for out-of-distribution generalization (ECCV 2024).
- Proposed an acyclicity-constraint-free causal discovery approach over an equivalence space of causal graphs (ICPR 2024 Oral; ICML 2021).
Causal Representation LearningDomain GeneralizationBayesian Inference
Graduate Research Assistant
May 2019 – Dec 2024Rensselaer Polytechnic Institute
- Built a unified framework combining causal discovery and representation learning for controllable and counterfactual image generation (paper under review).
- Investigated causal discovery under heteroscedastic noise, deriving implementable identifiability conditions and a two-phase iterative algorithm (AAAI 2024).
Causal DiscoveryRepresentation LearningImage Generation