Hugging Face's policy chief says AI model evaluations still lack reproducibility and standardization
Irene Solaiman, chief policy officer at Hugging Face, said during a discussion at TechCrunch Disrupt that the methods behind AI model evaluations are still maturing, arguing that the industry needs a more consistent approach before benchmarking results can be relied on.
Solaiman said trust in evaluation results depends heavily on reproducibility and on whether outsiders have the level of access needed to the model itself — a key argument for open weights. She noted that reproducibility has long been one of the stickier aspects of benchmarking, and that the industry still lacks meaningful ways to compare evaluations against one another because they are an evolving science. Evaluations start to become ineffective relatively quickly, she added, and the emerging embedded evaluator ecosystem needs far more standardization around who qualifies as an effective evaluator for different tasks.
Embedded evaluators — third-party observers placed inside AI organizations — have gained attention in recent months as AI labs respond to a spate of safety incidents and push for third-party oversight rather than active government regulation. Doubts remain about how neutral and effective such evaluators will be. Anthropic's selection of Accenture as its first embedded evaluator drew criticism from observers who questioned the consulting firm's qualifications.
Solaiman also said evaluations are expanding from models to the broader system around them, including the harness and the agent. The harness, or the tool used to access a model, has a significant impact on performance, and a weaker model with a better harness could outperform a frontier model. Agentic evals are really hard and really expensive, she said; there are more evaluations for models than for harnesses, and better evaluations are being developed for the holistic agent.
She concluded that the evaluation space remains fragmented, and if benchmarking is to remain useful to industry leaders, a more uniform, industry-wide approach may be necessary.
