Speaker
Description
Agentic AI systems are increasingly deployed in scientific workflows, yet existing evaluations primarily measure task completion and provide limited insight into scientific reliability. We argue that such metrics are insufficient for assessing autonomous systems in research settings, where plausible but incorrect results may be more dangerous than overt failures.
We present a structured evaluation framework for scientific AI agents and apply it to CMBAgent across eighteen astrophysical tasks spanning tool-grounded computation, Bayesian inference, and multi-step research workflows. The framework combines execution success, numerical fidelity, parameter recovery, physical plausibility, and failure transparency to assess both performance and scientific validity.
Our analysis reveals a consistent pattern across workflow paradigms. In the One-Shot setting, domain-specific context retrieval drives a ~6x performance improvement (Final Score 0.85 vs. ~0 without context); the primary failure mode without context is silent wrong computation — syntactically valid code producing plausible but numerically incorrect results. In the Deep Research setting, under-constrained inference tasks fail systematically: parameter degeneracies go undetected, yielding physically inconsistent posteriors reported as valid results. Compositional workflows exhibit inter-trial inconsistency and systematic bias without self-diagnosis. Across all tasks, failure transparency is the weakest dimension: the agent never proactively flags known pathologies in its own outputs.
These findings highlight a critical challenge for physics and astronomy: the dominant failure mode of agentic AI is not crashing or refusing a task, but confidently producing incorrect scientific conclusions without warning. We argue that reliability, physical consistency, and error detection should be treated as first-class evaluation objectives.