INTENT-AS-A-TOOL Makes it Easy to Track Agentic Misalignment

TL;DR - INTENT-AS-A-TOOL gives autonomous LLM agents a dedicated channel for signaling behavioral intent, enabling fine-grained tracking of agentic misalignment during reasoning. It could support earlier online intervention before harmful actions occur.

• Harmful actions under goal conflicts are often preceded by detectable intent signals in chain-of-thought reasoning.
• Intent-targeted tools let models explicitly indicate commitment toward a behavior during generation.
• Tool-call probabilities provide judge-free, dense trajectories of behavioral tendencies, unlike coarse post-hoc chain-of-thought labels.
• The method complements chain-of-thought monitoring and helps identify critical reasoning steps for intervention.