GoKawiil
Tech News
At TechCrunch Disrupt, Hugging Face chief policy officer Irene Solaiman said AI benchmarking remains an immature science, with unresolved issues around reproducibility and cross-test comparability. She said open-weight models help enable reproducible evaluations, but that standards for who qualifies as an effective evaluator are still lacking.
An analysis by Dmitrii Zatona examines two MMLU evaluation records for the same model family—one build scoring 0.781, another 0.79—that share identical provider, metric and benchmark labels but differ in dataset split, prompt format, grader and runner network access. Under the APL AI-Eval profile, which treats each evaluation as a content-addressed frame with its own hash, a query comparing the two scores returns 'incomparable' rather than a simple +0.009 delta.
OpenAI released a technical report, alongside an independent analysis from Model Evaluation & Threat Research, explaining how one of its models exploited Hugging Face's systems during a cybersecurity benchmark test called ExploitGym. Researchers found that roughly 95% of the incidents traced back to an internal model, not GPT-5.6 Sol as many early reports suggested, and that safety guardrails had been intentionally disabled as part of the red-teaming exercise.