Posts

Held at the ACM Conference on AI and Agentic Systems (CAIS), the RLEval: Methods and Reinforcement Learning Environments for Evaluating AI Agents workshop is believed to be the first-ever research venue for questions of AI agent evaluation and benchmarking. Last week, Brown CS PhD student Victor Ojewale and his advisor, Brown CS faculty member Suresh Venkatasubramanian (also Professor of Data Science, Deputy Director of the Data Science Institute, and Director of the Center for Technological Responsibility, Reimagination, and Redesign) received the event’s Best Paper Award.

“Most benchmarks for AI agents,” the authors explain, “usually only ask one question, did the agent complete the task? And what they don't ask is whether the agent should have completed the task in the first place. Our paper, ‘What Benchmarks Don’t Measure: The Case for Evaluating Abstention Competence in Autonomous Agents’, argues that this creates a blind spot, where agents learn to act even when they are missing critical information or haven't been given proper authorization to proceed.”

“We call this compliance bias, and we built a framework to measure it, including three metrics that capture whether an agent is both safe and genuinely useful. Testing across seven model families, we found the bias shows up in two opposite ways such that some models act too readily, others refuse too much, and the only thing that resolved both at the same time were checks that are installed outside the prompt interface we built as part of the work.”

Victor’s research focuses on algorithmic accountability, AI auditing, and evaluation infrastructure for large language models and autonomous agentic systems. His goal is to build rigorous, context-aware evaluation methods that make model behavior measurable, comparable, and auditable in real-world use. His recent honors include an IEEE SATML Distinguished Paper Award.

Suresh’s current research interests lie in algorithmic fairness, and more generally the impact of automated decision-making systems in society. His recent accomplishments include co-founding Brown’s AI Research Institute on Interaction for AI Assistants (ARIA), co-Chairing the Association for Computing Machinery’s US Technology Policy Committee’s Subcommittee on AI and Algorithms, and founding the Center for Technological Responsibility, Reimagination, and Redesign.

For more information, click the link that follows to contact Brown CS Communications Manager Jesse C. Polhemus.