P. Kamalaruban

Activation steering debiasing methods reduce LLM bias not by correcting underlying preferences, but by making the model less confident and more likely to abstain from answering.

FairTPT not only improves fairness in vision-language models but also prevents catastrophic forgetting, setting a new standard for test-time adaptation in AI.

Even with corrupted human feedback, surprisingly tight guarantees for multi-agent reinforcement learning are possible.

Forget perplexity: DMAP offers a mathematically grounded, model-agnostic representation of text that unlocks new insights into generation quality, machine-generated text detection, and forensic analysis of synthetic data influence.

RLHF and DPO are surprisingly vulnerable to data poisoning, with even a small number of carefully crafted preferences capable of steering the learned policy towards a desired (potentially harmful) target.