Hugging Face policy chief says AI evaluators need greater standardization
The brief
Irene Solaiman, Hugging Face’s chief policy officer, said AI evaluations remain an evolving science and are difficult to reproduce and compare consistently. She also warned that embedded third-party evaluators may not provide an easy solution, particularly as testing expands from models to the harnesses and agents built around them.
Key points
- Solaiman said evaluation results are more trustworthy when they can be reproduced and researchers have sufficient access to the model, including through open weights.
- Current evaluations are difficult to compare, can become ineffective quickly, and lack a sufficiently developed evaluator ecosystem.
- Embedded evaluators placed inside AI companies need clearer qualifications and greater standardization to assess different tasks effectively.
- Agentic evaluations are particularly difficult and expensive, while evaluations of model harnesses remain less developed than model testing.
- Anthropic’s hiring of Accenture as its first evaluator drew criticism from some observers who questioned the firm’s suitability.
Read the source report
TechCrunch AI ↗Full summary
AI labs rely heavily on benchmarks to compare models, but the methods behind those evaluations remain complicated and are still developing. Irene Solaiman, Hugging Face’s chief policy officer, said at TechCrunch Disrupt that the industry needs to develop a more consistent science before it can depend on evaluations as reliable measures of performance.
Solaiman identified reproducibility as a central issue in determining whether evaluation results can be trusted. She said access to the model itself, including access enabled by open weights, is important for reproducing results. She also noted that evaluations are difficult to compare with one another and can lose effectiveness relatively quickly.
The discussion also addressed embedded evaluators, in which third-party observers work inside AI organizations to assess safety and performance. AI labs have promoted this approach amid safety concerns, but questions remain about the evaluators’ neutrality and qualifications; Anthropic’s choice of Accenture as its first evaluator, for example, prompted criticism from some observers.
Evaluations are increasingly examining complete AI systems rather than models alone, including the harnesses used to access them and autonomous agents. Solaiman said a stronger harness can allow a weaker model to outperform a frontier model, while agentic evaluations are difficult and expensive; she argued that the industry needs more evaluation of harnesses and holistic agents alongside greater standardization overall.