<ul data-eligibleForWebStory="false"><li>Gradient-based optimization in deep learning raises privacy and security concerns due to data poisoning attacks and overfitting risks.</li><li>Black box optimization methods offer an alternative by treating the model as an opaque function, but face challenges in scalability and computational costs, especially in large language models (LLMs).</li><li>A new method called BBoxER is introduced for LLM post-training, inducing an information bottleneck via implicit compression of training data.</li><li>BBoxER provides theoretical bounds on generalization, privacy, data poisoning attacks, and robustness to extraction attacks, demonstrating promising results in experiments with LLMs.</li></ul>

Tuning without Peeking: Provable Privacy and Generalization Bounds for LLM Post-Training

Discover more