Skip to content

GateWatch: Low-overhead LLM monitoring with Boolean circuits

GateWatch: Low-Overhead LLM Monitoring with Boolean Circuits

arXiv: 其他 AI 安全论文· arXiv:2611.24619· Firas Spanioli, Yaniv David·Indexed by this site · 13 hr ago· Originally publishedAlignment· Its place in the scan record →
AI summary

The researchers propose GateWatch, a compact interpretable classifier that expresses the entire decision rule as explicit Boolean logic over a small number of binarized SAE features, for monitoring LLMs and LLM-based agents to detect unsafe behavior. The paper explores three methods for building these circuits under a given gate budget, and accompanies them with the GateWatchServe system, which computes only the selected SAE features, streams their binary values across layers at inference time, and evaluates multiple classifiers in parallel using compact lookup tables. On 24 tasks with Gemma-2-9B, GateWatch reaches an average test accuracy of 80.49% using only 3.8 SAE features and 3.5 logic gates per probe on average, 1.17 percentage points below a full 16K-feature SAE probe, ties or outperforms on 10 tasks, and reduces probe memory by 99.98%.

Read originalarxiv.org