UK AISI publishes reproducible evaluation results through EvalEval
The UK AI Security Institute (AISI) is using EvalEval’s infrastructure to publish evaluation methods and results in a more reproducible format. The collaboration, announced on September 22, 2026, follows joint research begun at a workshop held alongside NeurIPS 2025. Feedback from AISI also contributed to the development of the Every Eval Ever (EEE) schema.
Through Evaluation Cards, AISI is releasing verified results, configuration details and relevant context for five benchmarks from the main experiment of its paper How Inference Compute Shapes Frontier LLM Evaluation: HealthBench, FrontierMath, Humanity's Last Exam, SWE-Bench Pro and Terminal-Bench 2.0. The results cover Claude Opus 4, Claude Opus 4.5, Claude Opus 4.6, GPT-5, GPT-5.2 and GPT-5.4. Two related cyber evaluations, Cyber CTFs and The Last Ones, are included with a different, partly overlapping model set.
The organisations say open setup information can help researchers compare studies and examine how evaluation protocols and inference-time compute affect reported performance. EvalEval’s EEE schema and Evaluation Cards are intended to standardise these records across the evaluation ecosystem.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.