Detecting High-Stakes Interactions with Activation Probes
This addresses monitoring for safe LLM deployment, though it appears incremental as it builds on existing probe methods.
The paper tackles the problem of detecting high-stakes interactions in LLMs using activation probes, finding they generalize robustly to real-world data with performance comparable to LLM monitors while offering six orders-of-magnitude computational savings.
Monitoring is an important aspect of safely deploying Large Language Models (LLMs). This paper examines activation probes for detecting "high-stakes" interactions -- where the text indicates that the interaction might lead to significant harm -- as a critical, yet underexplored, target for such monitoring. We evaluate several probe architectures trained on synthetic data, and find them to exhibit robust generalization to diverse, out-of-distribution, real-world data. Probes' performance is comparable to that of prompted or finetuned medium-sized LLM monitors, while offering computational savings of six orders-of-magnitude. Our experiments also highlight the potential of building resource-aware hierarchical monitoring systems, where probes serve as an efficient initial filter and flag cases for more expensive downstream analysis. We release our novel synthetic dataset and codebase to encourage further study.