Amit Saha

Current Work

I'm actively working on the following projects:
  1. Privacy amplificiation via composite subsampling schemes for learning from structured data
  2. Private optimizers with preconditioning (e.g. Muon) in high dimension
  3. Memorization and generalization in diffusion models
  4. Privacy amplification by iteration with imperfect sampling oracles, applied to convex optimization
  5. Equivalences between notions of algorithmic stability in smoothed online learning
I have previously worked on the following projects:
  1. Frameworks & provable guarantees for algorithmic copyright protection
  2. Meta-evaluations of evaluation practices for AI safety

Conference Papers and Preprints

  1. Saha, A., Huang, Y., Chien, E., & Li, P. "The Access-Similarity Lens: An Operational Copyright Framework for Generative Models." Under submission at NeurIPS 2026; accepted to Theory and Practice of Differential Privacy (TPDP) 2026. May 2026 *
  2. Loehr, A., Khire, I., Huang, Y., Sharma, D., Saha, A., Hao, Y., & Manheim, D. "Systematic Review of Recommended Practices for the Design, Conduct, and Reporting of AI Evaluations through 2025." EvalEval @ ACL 2026, June 2026.
  3. Govil, S., Rodgers, J. P., Chou, Y.-T., Miao, S., Saha, A., Anand, A., Lieret, K., DeZoort, G., Liu, M., Duarte, J., Li, P., & Hsu, S.-C. "Locality-Sensitive Hashing-Based Efficient Point Transformer for Charged Particle Reconstruction", NeurIPS 2025 Workshop on Machine Learning and the Physical Sciences, Dec 2025.