Our AAAI 2026 poster on private in-context learning: LLMs can use sensitive records as in-context examples, with formal differential privacy guarantees and no fine-tuning.
Poster: https://doi.org/10.48448/8as1-z797
Paper: https://arxiv.org/abs/2509.13625
Code: https://github.com/bhusalb/privacy-preserving-icl
1/ Is your Gaussian mechanism actually (ε,δ)-DP? DP bugs are common and hard to spot by hand.
Our CCS 2025 paper shows that checking this for loop-free programs with Gaussian and Laplace sampling is almost decidable: decidable for every δ outside a finite set.
1/ Can an LLM use sensitive records as in-context examples without leaking them? Our AAAI 2026 paper: yes, with formal differential privacy guarantees and no fine-tuning.
Paper: https://arxiv.org/abs/2509.13625
Code: https://github.com/bhusalb/privacy-preserving-icl
1/ Muon is quickly becoming a go-to optimizer for training LLMs. What happens when you train with it under differential privacy?
New preprint with collaborators at Los Alamos: Inner Momentum for Differentially Private Muon