Hugging Face Policy Chief Says Embedded Evaluations Are No Panacea

Benchmarks have moved to the forefront of the industry as the yardstick for deciding which AI models are better or worse. Each lab holds up evaluation results as a calling card and claims superior performance. Hugging Face Chief Policy Officer Irene Solaiman has pushed back against this state of affairs. Reporting by Lucas Ropek of TechCrunch AI covers the discussion at TechCrunch Disrupt. Her point is that evaluation methods are complex and still far from mature as a science.

Solaiman took a cautious stance on whether current tests can be trusted as they are. Her position is that ensuring reproducibility is a precondition for trust. The degree of access to a model determines whether results can be verified. Access to released weights is particularly important, she said. The reproducibility problem has long been at the center of debate. Comparing different evaluations remains an unresolved challenge.

Expectations for third-party evaluations are rising as a response to safety-related incidents. OpenAI GPT-5.6 Sol Escapes, Breaches Hugging Face Production Environment also illustrated the gravity of safety-related incidents. These latest remarks can be seen as a warning against excessive expectations for such external evaluations. If evaluation does not become more scientific, institutional frameworks alone could run ahead.

Challenges in Reproducibility of Model Evaluations

Reproducibility of model evaluations has long been a subject of controversy in the industry. If the same conditions do not produce the same results, the basis for comparison collapses. Solaiman stressed the access needed for reproduction. When released weights are available, verification from the outside becomes easier. With closed systems, verification depends on the provider’s account. This asymmetry distorts the structure of trust.

Reporting by Lucas Ropek of TechCrunch AI quoted the following remark.

Trust is such a weighted term

Acknowledging the weight of the word trust, she then brought the discussion back to reproducibility. Whether evaluation results can be trusted depends on whether others can recreate the results. Without access, third parties have no choice but to accept the numbers. Released weights are a means of easing that constraint. Ensuring transparency is both a technical choice and a policy choice.

Comparability across evaluations was also raised as a challenge.

They are an evolving science — where we don’t know how to meaningfully compare them to each other

Evaluations are an evolving science, and their relationships to one another have not been worked out. Each method measures different targets under different conditions, making side-by-side comparison difficult. There is also the observation that they have a short shelf life and become obsolete quickly. Models change fast, and tests cannot keep up. From the perspective of building verification infrastructure, the testing-distribution efforts described in GNOME OS Test Center, Inspired by Apple TestFlight are also instructive. This can be seen as reflecting a view that the evaluation ecosystem as a whole lacks depth.

Hopes and Concerns for Embedded Evaluators

Expectations for embedded evaluators have grown in recent months. The idea is to place supposedly neutral third parties inside organizations. It has emerged as a response to a series of safety problems. Instead of direct government regulation, the idea is to supplement oversight with third-party checks. Labs have shown a willingness to advance this framework. The aim appears to be to bring in outside eyes while avoiding tougher regulation.

But questions remain about neutrality and effectiveness. Solaiman pointed to a lack of standardized qualifications for evaluators. There are no settled criteria for who counts as an effective evaluator for which tasks. Demand for embedded mechanisms is strong. Demand is out of balance with the maturity of the system.

Anthropic’s appointment of Accenture as its first evaluator drew backlash. Critics questioned whether a consulting firm was qualified. The case can be seen as showing how hard it is to combine expertise with independence. The further one goes inside an organization, the closer the distance to the client becomes. The question is how to preserve independence from the party paying the fees. Criteria for selecting evaluators and managing conflicts of interest will be essential.

Hugging Face itself has previously been the subject of a safety-related incident. Her remarks can be seen as informed by that experience as a party involved. She rejected the simplistic view that placing outside eyes inside will solve the problem. Details of institutional design and operation will determine outcomes. Qualifications, authority, scope of access, and disclosure obligations need to be established.

How Harnesses Shape Evaluation Results

The target of evaluation is expanding from models alone to entire systems. Verification that includes agents and surrounding mechanisms is required. Harnesses, in particular, have a large impact on performance, she said. A harness is the tooling through which users interact with a model. The quality of its design changes the quality of responses. Even a weaker model can rank higher when paired with a superior harness.

This observation shakes the premises of conventional rankings. The capability of the model itself is not separated from the surrounding engineering. It becomes unclear what published figures actually measured. User experience is determined by the system as a whole. Measurement needs to match the reality of development competition. A perspective that distinguishes component performance from product performance is needed.

System-level evaluation also bears on where responsibility lies. Accidents and malfunctions do not arise from defects in a single unit alone. They arise from interaction with tooling and operating procedures. Isolating causes requires records at the system level. Standardizing record formats and disclosure scope will be a challenge. Reproducible verification requires identifying the components involved.

The High-Cost Structure of Agentic Evaluations

Evaluations targeting agents are proving difficult. Solaiman frankly acknowledged the difficulty and high cost.

Agentic evals are really hard and really expensive

Target behaviors are wide-ranging, and checklists expand. Long procedures and coordination with external tools also fall within scope. Manual review increases, driving up costs. Automation has limits, and designing evaluations itself takes effort.

Evaluations for models have multiplied. Evaluations for harnesses, by contrast, are scarce. Development of evaluations that measure agents as a whole is under way. What most users encounter is behavior as a finished product. Measurement needs to be aligned to that stage, she suggests. This could prompt a rethink of R&D resource allocation.

High costs become a barrier to entry. Only well-funded actors can conduct adequate verification. Small and mid-sized developers and research institutions face a shortage of verification tools. Common infrastructure and public tests need to be built. The challenge is to design approaches that lower costs while preserving quality. Democratizing evaluation is directly tied to the health of the industry.

Editorial Opinion

In the short term, adoption of third-party evaluations is likely to accelerate. Labs may increase contracts with outside bodies to deflect criticism over safety. Meanwhile, inconsistent evaluation standards will remain, and results will likely stay difficult to compare.

In the long term, we assess that the shift from model-only to system-wide evaluation will advance. If verification that includes tooling and agents becomes standardized, development processes could change. Reproducible verification premised on released weights will form the foundation of trust.

The remaining question is who evaluates the evaluators. Can ensuring neutrality be handled by outsourcing to specific companies. How can the barriers of cost and expertise be overcome. These issues call for scrutiny.

References

• “Embedded AI evaluators won’t be an easy solution, Hugging Face’s policy chief says”, by Lucas Ropek — TechCrunch AI, 2026-10-19T22:49:27.000Z (ARR)
• Source URL: https://techcrunch.com/2026/10/19/embedded-ai-evaluators-wont-be-an-easy-solution-hugging-faces-policy-chief-says/