From the source
From the source
From the source

From the source
Lightweight activation probes already help screen model activity in production monitoring systems, routing flagged behavior for further review. 1 We study the behavioral properties of such probes: what they flag, which parts of an interaction they respond to, and when they first raise an alert.
Our case study uses a small probe trained on frozen Kimi K3 activations using an LLM misalignment monitor’s labels. Here are some key observations from our study: Distinguishing violations from benign work: We show that the probe can distinguish monitor-flagged misalignment from benign behavior in contextually aligned tasks and rollouts.
Looking beyond final answers: Strong probe signals appear in reasoning, tool calls, and tool feedback, including plans and observations surrounding a violation. Overlap with monitor evidence and late alerts: The probe’s signals overlap with the monitor’s citations in many runs that both systems flag, even though we train on labels for whole trajectories.
However, some first alerts arrive only in later tool feedback or user-facing claims, potentially leaving too little time to intervene. …
commandline.microsoft.com