Embedded AI Testers Won't Fix Benchmarking on Their Own
Hugging Face chief policy officer Irene Solaiman said embedded AI evaluators will not be an easy fix for the problems with AI model benchmarking, speaking at TechCrunch Disrupt on October 19, TechCrunch reported.
Solaiman said there is strong demand for an ecosystem of embedded evaluators, meaning testing experts who work close to where models are built and used. She said qualifications for effective evaluators still need more standardization.
She described AI evaluation as an evolving science. It is still unclear how to compare one evaluation against another in a meaningful way.
Reproducibility, she said, comes down to access. Whether another group can rerun a test and get the same result depends on how much access it has to the model. She pointed to open weights, where a model's core parameters are shared publicly, as a way to make that checking possible.
She said evaluations for agentic systems, or AI that carries out multi-step tasks with tools, are hard and expensive. Evaluations for harnesses, the surrounding software of prompts, tools, memory and product logic, lag behind evaluations for models.
The broader context here will be familiar to practitioners. A model score alone says little once prompts, tools, memory, retrieval and product logic shape behavior. Choices about the harness can shift results as much as changes to the model. Agentic runs add long timelines, tool calls and randomness. That makes comparison across tests unreliable without shared methods, versioned test conditions, and disclosure of what system was actually tested.
In my view, Solaiman put the bottleneck in the right place. The limit is not interest in evaluation. Teams want continuous, embedded checks rather than one-off leaderboard runs, because static benchmarks saturate and leak and drift from deployed use. The limit is agreement on who counts as a qualified evaluator and what good practice requires. Without that, more evaluators produce more numbers, not more clarity. Standardization sounds unglamorous. It is the work that lets results add up across labs, vendors and auditors.
Looking at what this means for builders, the near-term agenda is practical. Treat harnesses as first-class test targets, not invisible glue, with their own regression checks, failure lists and misuse cases, separate from the base model card. Design evals to be repeatable across access levels, and record what access a result assumed. Open weights help here, because they let others rerun, probe and vary conditions rather than take a score on trust. The payoff is not perfect measurement. It is measurement that improves with use, and gives product teams, researchers and policymakers a firmer basis for deployment and risk decisions.
