Ching Lam CHOIReducing LLM bias and toxicity by constraining the reward model to be independent of sensitive attributes, conditional on unrestricted features.
Preprint 2026
Fairness Aware Reward Optimization
Related work
Preprint 2026
Tabular Pancakes: Cryptographic Watermarks Against Dataset Fabrication
Watermarking tabular data generators by sampling latents from the keyed Pancake distribution (hCLWE) instead of Gaussian, for detecting fraudulent data fabrication.
Preprint 2022
Universal Adversarial Directions
UAD is an adversarial examples game with a pure Nash equilibrium and better transferability.