A Vocabulary for Multi-Agent Automated Research Systems

The author proposes a new vocabulary for multi-agent research systems: Generative taste and Evaluative taste.

Generative taste refers to the ability to suggest novel research directions, while Evaluative taste refers to judging outcomes accurately.

The proposal is framed as a specification language rather than a benchmark, and it is supported by case studies.

Its goal is to separate idea generation from scoring quality so researchers can better diagnose failures in agent pipelines.