Tag
This paper analyzes the privacy-utility trade-off in LLM interactions, uncovering underlying mechanisms and introducing an intent-driven local protection framework with a lightweight model to enhance privacy while maintaining response utility.
DT-Guard introduces a reasoning-active training, reasoning-free inference paradigm for LLM safety guardrails, achieving strong F1 scores on safety benchmarks with a 4B backbone model, outperforming larger baselines.
An experiment on using Rust's type system as an AI coding specification, distilling practices from Epic Lora.