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

The paper introduces BBoxER, an evolutionary black-box optimization method for post-training large language models (LLMs) that avoids gradient exposure, addressing privacy and security concerns. BBoxER employs an information bottleneck and provides non-vacuous generalization bounds, demonstrating improved performance on reasoning benchmarks and robustness against membership inference and data poisoning attacks. This method offers a viable alternative for practitioners seeking to enhance LLM training in sensitive environments while ensuring strong theoretical guarantees.