Hugging Face policy chief says AI benchmarks need common standards

What a benchmark can establish

Solaiman said that asking whether a test can be trusted bundles together several questions. Can researchers reproduce its results? Do they have enough access to the model itself? Open weights can help with access, but they do not resolve the wider problem of comparing benchmarks.

The tests also lose effectiveness relatively quickly, she said, while the ecosystem needed to support rigorous evaluation is still underdeveloped. Researchers have yet to establish a meaningful way to compare tests against one another.

That makes standardisation of evaluators a practical concern, not just a question of terminology. Interest in embedded evaluators is growing, but the criteria for deciding who can assess different tasks well still need much more work.

Independence is part of the test

Embedded evaluators are independent observers placed inside AI organisations. Their appeal has grown in recent months as labs have faced safety incidents and concerns about potential harm from their technologies. Hugging Face was itself targeted in one such incident.

Labs are advocating third-party evaluation instead of active government regulation to identify and reduce model risks. But an outside evaluator is not automatically independent or effective. Anthropic’s decision to hire Accenture as its first evaluator drew criticism, including doubts about whether the consultancy had the right expertise.

The system matters as much as the model

Evaluation is expanding beyond the model to the system around it, including the wrapper through which a user interacts with the model and the agent that carries out tasks. Solaiman said a weaker model with a better wrapper can easily outperform a frontier model.

Agent evaluations are difficult and expensive, and there are more tests for models than for wrappers. Researchers are developing methods better suited to whole agents, the systems many users actually rely on.

I think that shift exposes a weakness in the usual benchmark story: a score can look like a verdict on a model when it also reflects the tools and setup around it. The harder question is who sets consistent standards for judging that full system—and whether labs’ chosen evaluators can meet them.