safety-detection

PolicyMem proposes a geometric policy memory for LLM governance, externalizing natural-language policies as reusable geometric objects to enable detect-rewrite-verify loops with state-of-the-art unsafe behavior detection.

This paper proposes TRACE, a trajectory-level safety detection method for long-horizon LLM agents that compresses full trajectory evidence into a latent state to better aggregate dispersed risk signals, achieving state-of-the-art accuracy on multiple benchmarks.