Hugging Face policy lead: embedded AI evaluators are hard to make reliable

Solaiman said evaluation credibility hinges on whether results can be reproduced and whether evaluators can get sufficient access to the model itself, which is a key value of open-weight models. She sees evaluation as an evolving science: the industry doesn't yet know how to meaningfully compare different tests, and methods go stale quickly. Demand for embedded evaluators is rising, but standards for who is qualified to evaluate which tasks need major standardization. Evaluation is expanding from the model itself to the whole system around it, including the harness users access models through and agents. Solaiman said a weaker model with a better harness can beat a frontier model; agent evaluations are hard and costly, and far fewer harness evaluations exist than model evaluations. For teams benchmarking open-weight models and enterprises relying on third-party evaluations for selection, the short term still requires building reproducible evaluation pipelines in-house; including harnesses and agents will cost notably more than running model tests alone.