Can AI model benchmarking be trusted? Not fully yet, according to Irene Solaiman, Hugging Face's chief policy officer. Speaking at a discussion at a technology conference, she said the methods behind these tests are complicated and still maturing, and that the field needs more consistent science before the industry can rely on them heavily.
AI labs have built their reputations on benchmarks meant to show, in a neutral and accurate way, whether one AI product is better than another. Solaiman said trust depends first on whether a result can be reproduced, meaning whether someone else can run the same evaluation and get the same answer. She also pointed to access to the model itself, and to open weights, which are the publicly released parameters of a model, as a major way to make that possible.
Why the results are hard to check
Reproducibility in AI model benchmarking has long been one of the difficult parts of the field, and it remains a subject of debate across the industry. Even when a single test is sound, there is a separate question of how it compares with other evaluations. Solaiman said the field does not yet know how to meaningfully compare them with each other.
She added that tests can become ineffective fairly quickly, and that the ecosystem of groups building and running them is not as strong as it needs to be. Her view was that the industry needs more standardised methods before scores can be treated as settled.
Harnesses, agents and embedded evaluators
Evaluations are moving beyond the model alone and now cover the whole system around it. That includes the harness, which is the tool a user works through to reach the model, and the agent that carries out tasks. Solaiman said a harness can have a large effect on performance, and that a weaker model with a better harness could outperform a stronger one.
She also said there are far fewer evaluations for harnesses than for models. "Agentic evals are really hard and really expensive," she said. Her remarks came as the idea of embedded evaluators, outside observers placed inside AI organisations, has gained attention over recent months. Solaiman said there is strong demand for such an ecosystem, but that the qualifications needed to judge who is an effective evaluator for which tasks require far more standardisation.
The push for these arrangements follows a run of safety incidents across the sector. Rather than active government regulation, labs have largely turned to third-party checks. Hugging Face was itself the target of one of those incidents. Anthropic's decision to hire Accenture as its first evaluator drew backlash, with some critics questioning whether the consulting firm was qualified for the job.
What comes next
Solaiman's comments suggest that the current evaluation landscape is still loosely organised. If AI model benchmarking is going to remain useful to industry leaders, she indicated, a more uniform, industry-wide approach may be needed so that everyone works from the same standards.
Aasan alfaaz mein
AI companies apne models ko dusre models se compare karne ke liye tests use karti hain, lekin Hugging Face ki chief policy officer Irene Solaiman kehti hain ke yeh tests abhi poori tarah bharosemand nahi. Nateeja dobara parakhna mushkil hai aur ek test ka doosre test se muqabla bhi aasan nahi. Is liye AI tool chunte waqt sirf score ko aakhri sach na samjhein.
