Misalignment signals in Kimi K3: What probes flag and where they peak
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
A study of lightweight activation probes trained on frozen Kimi K3 activations, using labels from an LLM misalignment monitor, examines what these probes flag during agent runs. The authors find the probe separates monitor-flagged violations from similar benign behavior and often fires on reasoning, tool calls, and tool feedback.