4 trigger types do not find the correct time to stop an AI agent
3 human annotators agreed on where to stop the agent at a rate above chance level by a small quantity.
Claimed, not confirmed
Researchers tested 4 types of intervention trigger that stop an AI agent. They used debugging traces from SWE-bench Verified and compared them with the stop points that humans marked. An intervention trigger that uses a threshold on modeled frustration was active on 39% to 83% of the actions. 3 human annotators agreed on where to stop at a rate above chance level by a small quantity.
Sources
Posted
