On the job market this year for research roles

Research in structured 3D vision and efficient, interpretable AI.

I work at the intersection of computer vision, 3D perception, efficient adaptation, and cognitive AI. My research asks how models can recover the structure humans use naturally: parts, scenes, geometry, causal organization, and reusable internal representations.

The core theme is analysis by synthesis: infer hidden structure by building and testing generative explanations. This leads to practical systems for 3D understanding, robust evaluation, and continual learning without brute-force scale.

3D part representation learning: part-centric semantics, object pose, reconstruction, and synthetic-to-real transfer.
Physics of AI and mechanistic interpretability: spectral structure, update geometry, model merging, and learning dynamics.
Efficient learning and inference: parameter-efficient adaptation for LLMs, vision-language models, diffusion models, and 3D models.
Interpretable and robust learning: failure-aware evaluation for large models and 3D vision systems under shift, occlusion, and hallucination.

Selected publications

All publications

2025 · Physics of AI / mechanistic interpretability

The Universal Weight Subspace Hypothesis

A Physics of AI / mechanistic interpretability study showing that backprop-trained neural networks tend to converge to similar low-rank weight subspaces across datasets and hyperparameters.

Project · arXiv · Code

2026 · 3D vision / part semantics

Name That Part: 3D Part Segmentation and Naming

A one-shot system for predicting complete 3D part decompositions and assigning semantic names.

Project · arXiv

2026 · 3D evaluation / hallucination

Can These Views Be One Scene?

Shows that learned 3D backbones can hallucinate dense geometry for unrelated scenes, repeated views, and random noise; introduces failure-aware consistency metrics.

Project · arXiv · Code · X post