Interactive Evaluation Requires a Design Science

In three linesPosition paper on interactive evaluation of LLMs. Models deployed as systems acting over time (tools, environments, agents) require evaluation paradigm distinct from static benchmarks. Authors propose taxonomy, design principles, and reporting standards to assess process, recoverability, coordination, robustness, and system-level performance.

Summary generated by Claude — human-verified