On October 28, 2026, researchers from Harvard, MIT, UMass Amherst, Northeastern and industry met in Boston for the second AES Boston working group. Participants work on benchmarks and RL environments for scientific and clinical work, statistics for evaluation, multi-agent safety, LLM judges, interpretability, data systems, personal research agents and AI for materials science.
We went around the room with introductions, then had three short talks with discussion after each. This post summarizes what we heard. We hope readers take away two points. First, generating tasks, environments, analyses and even papers with AI is now easy, and checking them is the bottleneck. Second, much of what we rely on to check agents (final scores, LLM judges, the model's own reasoning) is weaker than it looks, and this matters most in RL environments and in safety evaluation.
Four problems that came up repeatedly
Checking AI's work costs more than producing it
Groups that build benchmarks and environments from scientific papers or clinical data now use AI to generate tasks at scale, and the human quality check has become their most expensive step. Several participants had also found errors in AI-assisted papers that nobody caught. Nobody in the room had a good way for humans and AI to share the checking without losing trust.
Environments shape what agents learn and what we can measure
The same environment is often used both to train agents and to evaluate them. Participants discussed how realistic an environment needs to be, how to tell which part of it drives a result, and how training in it shapes behavior beyond the task. Some settings that matter, such as personal research agents and materials science, have no good environment yet.
Safety auditing cannot rely on reasoning traces
A model's stated reasoning often does not match its actions, and participants working on multi-agent safety see behaviors like deception and collusion emerge without being prompted. Auditing therefore has to look at actions, which is hard when nobody knows what the right action was. Participants agreed that safety auditing today relies heavily on intuition, and that we need environments built for auditing.
Neither the judge nor the human is a gold standard
LLM judges are usually validated against human labels, but experts disagree and make mistakes too. Several participants thought methods from statistics and the social sciences could help, and that agent evaluation rarely uses them.
Three short talks
The first talk asked whether agents should be specialized by model architecture as well as by prompts. The discussion focused on when specialization could beat general models that keep improving, and the room thought it fits best in scientific domains where regular agents fail.
The second talk asked when we can say AI has taken over part of a science, and how we might measure it, drawing on the recent debate in mathematics and on examples from epidemiology. The discussion focused on what evidence exists today and how quickly it goes out of date.
The third talk was on auditing agents for safety, including what a model's reasoning can reveal and how to audit actions in real deployments. The discussion focused on how safety monitoring is done in practice today.
What worried people
Participants worried that the work of checking AI output falls on a small number of experts, often unpaid, and that the evidence on whether AI helps experts is thin and goes out of date quickly. On safety, they worried that monitoring of deployed agents still runs mostly on intuition. Early-career participants from outside computer science also said fellowships, mentors and compute for this work are hard to find.
Open questions
A few questions came up that nobody in the room could answer:
- How much expert time does checking AI's work take, and when does AI actually save it?
- How should a judge be validated when no human rater is a gold standard?
- How do we design environments that are realistic and still auditable?
- How do we audit an agent for safety when its reasoning cannot be trusted?
What EvalScience did in October
The Boston working group now has two organizers from its own members, Abhinav Kumar (UMass Amherst) and Kaden Zheng (Harvard). A Bay Area gathering met in October, and a New York working group starts in December with Kyuseong Choi (Cornell Tech) and Meet Udeshi (NYU Tandon) as organizers. Mentored Projects, our mentorship program for people entering agent evaluation and safety evaluation, has 10 mentors and more than 40 mentees signed up, and Stephan Rabanser has joined as Research Lead. Hanna Wallach (Microsoft Research), Kathryn Hume (Vector Institute), Eugene Wu (Columbia), Csaba Szepesvári and Issa Dahabreh (Harvard Chan) have joined our advisory board. The Agent Evaluation Science Symposium in New York on November 20 is supported by Schmidt Sciences, Jane Street and Fleet AI. We are co-organizing the AAAI-27 Bridge Program "From Agent Benchmarks to Deployment Evidence" in Montréal in February 2027, and several workshop proposals are in progress.
Next steps
The Boston working group meets monthly, with a shared Slack between sessions. Participants present their own work, which we feature on evalscience.org, and we invite guest speakers, including researchers from our Bay Area group. Safety evaluation and RL environment design will be a large part of the coming sessions. This semester, the group's shared effort is AES 2027 Spring, our Boston summit organized with Microsoft. We are also connecting people who have scientific questions and data with people who have evaluation methods, and starting a mentorship program with an open call, where mentors (including participants from this session) can bring their own projects. The next session is on Wednesday, December 2.
AES runs similar groups in the Bay Area and, starting in December, New York, as well as the Agent Evaluation Science Symposium in New York on November 20.
If you work on any of these problems, we would like to hear from you at contact@evalscience.org. The group runs on volunteer time, so we would especially like to hear from institutions that can host sessions or provide compute for participants' projects.