AI Laboratory has always bet its reputation on model benchmark testing – this supposedly neutral and accurate evaluation system is designed to measure whether one AI product is superior to another.
However, Eileen Solomon ( Irene Solaiman ), the Chief Policy Officer of Hugging Face, stated last week in a discussion at TechCrunch Disrupt that the methods behind these assessments are quite complex and are still evolving. If the industry is to rely on these tests in the future, it will be necessary to develop them into a more consistent science.
Considering this, can we trust these tests that are currently being released?
"Trust is a very important concept," said Solomon, "If we are to trust the results, a large part of that depends on whether you can reproduce the results of an assessment once again. Do you have the necessary access rights to the model itself? That's where open weights play a significant role."
Reproducibility in model evaluation has long been a challenging aspect of benchmark testing and remains a topic of intense debate within the industry.
In addition to whether the individual test itself is effective, there is another additional question: how does this test compare to other assessments?
Solomon said, "This is a science that is constantly evolving—we still don't know how to compare them meaningfully with each other."
She also stated, "They will relatively quickly become inefficient, and we haven't established the kind of prosperous ecosystem that is needed. Now everyone is eager to build an embedded evaluator ecosystem, but there is still a need for more standardization to determine what qualifications are required for someone to be an effective evaluator in what tasks."
In the past few months, the concept of "embedded evaluators" has gradually become popular. The approach involves placing so-called "neutral" third-party observers within the organization of AI. As the AI laboratory has encountered a series of security incidents and concerns about the potential hazards of this technology have grown from outside, this approach has received even more attention. Hugging Face was also the target of one of these incidents.
AI Laboratories are promoting this third-party approach as an alternative to proactive government regulation in identifying and mitigating model security issues. However, there is still skepticism from the outside world about how neutral and effective these assessors can truly be. Anthropic The decision to hire Accenture (Accenture) as its first assessor has sparked quite a backlash; some critics have mocked that the qualifications of this consulting firm to undertake this task are in themselves questionable.
Today, evaluation is shifting from simply testing the model itself to testing the entire system surrounding the model – including harness and agent. Solomon stated that what is known as harness, which is the tool used by users to access the model, has a significant impact on the model's performance; a weaker model, when paired with a better harness, is quite capable of easily outperforming a more advanced model.
“Agentic evals is really difficult, and also very expensive,” said Solomon, “At present, our focus is more on the evaluation of models. …… There isn’t as much evaluation for harness, and we are also developing better evaluation methods for the overall agent – because it is at this level that many people truly obtain value.”
In other words: to some extent, the current evaluation fields are still operating in a state of isolation. If benchmark testing is to continue to be regarded by industry leaders as a useful tool, then there may be a need for a more unified approach that covers the entire industry, so that everyone is on the same page.
