evaluator

Beyond the Search Box: Defining Real Research in AutoResearch

When an agent optimizes a score inside a research space that humans already defined, are we evaluating discovery or refinement? The distinction matters. Searching thousands of solution variants is useful, but it is not the same as asking whether the problem itself should change. Real research judgment questions the framing, tests transferability, and recognizes local optima. That is why the FLEET approach to reward-aware search is worth reading, it tackles the optimization side, yet the deeper question remains open.