INTENT-AS-A-TOOL Makes it Easy to Track Agentic Misalignment
The paper proposes tracking misalignment by giving agents an explicit “intent” channel and watching when they use it.
The authors argue that harmful agent actions are often preceded by detectable intent signals in chain-of-thought reasoning. Their method, INTENT-AS-A-TOOL, adds intent-targeted tools so the model can express commitment to a target behavior during generation. The probability of calling those tools becomes a fine-grained signal, without relying on a separate judge. They say it complements CoT monitoring and can identify steps where online intervention may be possible. ArXiv · AI/CL/LG's note
The authors argue that harmful agent actions are often preceded by detectable intent signals in chain-of-thought reasoning. Their method, INTENT-AS-A-TOOL, adds intent-targeted tools so the model can express commitment to a target behavior during generation. The probability of calling those tools becomes a fine-grained signal, without relying on a separate judge. They say it complements CoT monitoring and can identify steps where online intervention may be possible. ArXiv · AI/CL/LG's note
score 5