Justin Kang
I am a PhD candidate at UC Berkeley (EECS), affiliated with BAIR and advised by Prof. Kannan Ramchandran. My research develops efficient algorithms for ML interpretability and attribution — explaining which input features, training data, and interactions drive model predictions in LLMs and other large-scale models.
Research Highlights
- Interpretability & Attribution: I build scalable tools (SPEX, ProxySPEX) that identify important feature interactions in LLMs, achieving up to 20% better faithfulness than prior methods like SHAP, and scaling to 1000+ input features. Check out the shapiq library to try it out!
- Signal Processing → ML: I bring a strong signal processing and information theory perspective to ML problems, which leads to unique algorithmic solutions — including sparse Möbius/Fourier transforms for efficient model explanation.
- Faithfulness of Explanations: I recently led work on evaluating whether LLM self-explanations are faithful to actual model behavior in collaboration with Noah Siegel from Google Deepmind.
- Award-Winning Research: My work on scheduling in massive random access networks won the 2024 IEEE ComSoc & IT Society Joint Paper Award. Joint Paper Award
Recent News
- Talk at ITA 2026 graduation day — won the sea award. (slides)
- New preprints: Adaptive Sparse Möbius Transforms, A Positive Case for Faithfulness, and An Odd Estimator for Shapley Values.
- Papers ProxySPEX (Spotlight) and SHAP-Zero (Poster) accepted to NeurIPS 2025.
- SPEX: Scaling Feature Interaction Explanations for LLMs accepted to ICML 2025.
- Presented research on interpreting LLMs at DEVCOM Army Research Lab.
Older Announcements
- Joining Bosch AI Research for summer 2025, working on autolabeling and data filtering with Suraj Srinivas, Jorge Piazentin Ono, and Jared Evans.
- Learning to Understand: Identifying Interactions via the Mobius Transform accepted to NeurIPS 2024.
- Interviewed for a Built In article on Federated Learning.
- My paper won the 2024 IEEE Communication Society and Information Theory Society Joint Paper Award.
- Joined Google for summer 2024 as a Student Researcher in the Cloud Platforms Systems Research Group (SRG).
Selected Papers
A Positive Case for Faithfulness: LLM Self-Explanations Help Predict Model Behavior. Mayne, H.(e), Kang, J.S.(e), Gould, D., Ramchandran, K., Mahdi, A., Siegel, N.Y. Preprint 2026 paper
An Odd Estimator for Shapley Values. Fumagalli, F., Butler, L., Kang, J.S., Ramchandran, K., Witter, R.T. Preprint 2026 paper
ProxySPEX: Inference-Efficient Interpretability via Sparse Feature Interactions in LLMs. Butler, L(e), Agarwal, A.(e), Kang, J.(e), Erginbas Y.E., Ramchandran, K., Yu, B. NeurIPS 2025 Spotlight paper
SHAP-Zero explains biological sequence models with near-zero marginal cost for future queries. Tsui, D, Musharaf, A, Erginbas, Y.E., Kang, J.S., Aghazadeh. NeurIPS 2025 paper
SPEX: Scaling Feature Interaction Explanations for LLMs. Kang, J.S(e)., Butler, L.(e), Agarwal, A.(e), Erginbas Y.E., Pedarsani, R., Ramchandran, K., Yu, B. ICML 2025 paper
Learning to Understand: Identifying Interactions via the Mobius Transform. Kang, J.S., Erginbas, Y.E., Butler, L., Pedarsani, R., Ramchandran, K. (2024) NeurIPS 2024 paper · video
The Fair Value of Data Under Heterogeneous Privacy Constraints in Federated Learning. Justin Singh Kang, Ramtin Pedarsani and Kannan Ramchandran TMLR 2024 paper · video
Industry Experience
- Google — Student Researcher, Cloud Platforms Systems Research Group (Summer 2024)
- Bosch AI Research — Research Intern, working on autolabeling and data filtering (Summer 2025)
- Intel — Non-Volatile Memory Solutions Group, Storage Systems Research Intern (previously)
Education
- PhD in EECS, UC Berkeley (in progress)
- M.A.Sc. in ECE, University of Toronto — advised by Prof. Wei Yu
- B.A.Sc. in Engineering Physics, University of British Columbia