Materials ยท corvid/revised/cover-letter.txt
Jordan Hale jordanhale.example | [ASK JORDAN: email] | [ASK JORDAN: phone] | [ASK JORDAN: city and state] [ASK JORDAN: submission date] Corvid Labs hiring team Re: Research Scientist, Agent Evaluation Dear Corvid Labs hiring team, In your March engineering blog post, you wrote that your automated graders and your human raters often disagree on long tasks, mostly where an agent reached the right end state by a route a person would reject, and that you do not yet know which side is in error. I have worked on a smaller problem with the same structure. Since 2024 I have built and funded Relay, a testbed for studying how information changes as it passes along chains of language-model agents. To measure those changes at scale, I used language-model judges to label each message, for example whether a fact had been dropped, or a qualification such as a hedge or caveat. Before trusting the judges, a research assistant and I labeled a set of messages by hand, and I compared the judges with our labels. The judges agreed with us well on dropped facts and poorly on dropped qualifications. The choice between a cheap automated judge and slow human labeling therefore had a different answer for each property, and one overall agreement figure would have hidden the property where the judge was weak. [ASK JORDAN: what you did with the qualification labels after the judges agreed poorly.] Before the first full run I also found that the route through which the agents reached their models was adding hidden context, text the design did not include, to every call, which would have contaminated every condition. A paper reporting the Relay experiments is under review at Management Science, in its first round. Relay messages are text, not multi-step computer-use trajectories, so the parallel is one of structure. My judges were reliable on the blunt property and unreliable on the subtle one, and your disagreements fall along a similar line, between end state and route. In Relay I treated the hand labels as the reference. For you, whether the raters can serve as the reference is part of the open question. That is where I would start. First, settle whether the route is part of success, since graders and raters cannot be compared until both are scoring the same thing. Then measure how well raters agree with one another on the disputed trajectories, score end state and route separately, and only then sort the disagreements into grader errors and rater errors. [ASK JORDAN: whether you measured agreement between yourself and the research assistant in Relay. If so, say so here.] Your head of research has described the main risk as a score that measures the wrong thing. I have met that problem with people as respondents. At Kestrel Media, the main item in our satisfaction survey tracked recent service outages more closely than viewers' satisfaction with the catalog. I rebuilt the measure and checked it against which viewers later canceled. My work with human participants comes from laboratory experiments on group decisions during my postdoc, where I also managed the participant pool. I have not built reinforcement-learning environments or trained models. Sincerely, Jordan Hale