AI Model Benchmarking for Nonproliferation Applications: Steel Thread Benchmarking Task Force Technical Report (Rev. 2)
Steel Thread is a NA-22 venture that seeks to build trustworthy, reliable AI models that can be used in a wide variety of nonproliferation tasks. A key aspect of building these models is developing appropriate benchmarks and evaluation methods, which will enable the venture to identify and adapt models to provide the most value in the nonproliferation domain. Benchmarks must be relevant to key tasks in this domain, such as question answering, information retrieval, document summarization and classification, consensus analysis, and image and data analysis. This report 1) provides an overview of benchmark design, evaluation, and challenges; 2) reviews a variety of open benchmarks, with a focus on language models and tasks; and 3) identifies benchmarks that are most relevant to Steel Thread. This report is intended to serve as a basis for further efforts to classify and evaluate benchmarks and their correlation with success on nonproliferation-specific tasks. The Steel Thread venture has defined benchmarks to be a particular combination of a dataset (or datasets) and a metric (or metrics) conceptualized as representing one or more specific tasks or sets of abilities for a specific modality. It is adopted by a research community as a shared framework for comparing methods.1 It includes 1) Data: Labeled (a designated subset not used for training, which could be all the data), 2) Metric: A way to quantify performance, 3) Task/Ability: What the benchmark is testing, 4) Protocol: A structured and repeatable evaluation process, 5) Baseline/Reference Model: For comparison; could be statistical, rule-based, SME-derived, or another model, and 6) Maintenance Plan: to update with new information over time; important for long-term utility. For further clarity, the definition includes what a benchmark, in this context, is not. It is not a corpus of training data, specific to a model (it is intended to apply to a range of models), a universal evaluation of performance, a guarantee that the ‘top’ model on the leaderboard will be the best fit for every specific use case, an all-encompassing proof of a model’s universal quality, nor is it a one-size-fits-all measure of success. It does not cover every real-world constraint (like operational, ethical, or cost considerations), a systems integration test, or a unit test. This definition was inspired by and resulted from discussions within the Steel Thread Benchmarking Task Force. This group was formed to define what we would mean as a benchmark within Steel Thread but persisted as the need to develop a thorough understanding of the large and expanding existing benchmarking space. This technical report is a result of the group’s divide and conquer approach to exploring this space. The release of benchmarks might not be progressing as quickly as model development, but it is moving very fast, as many benchmarks quickly become saturated, when state-of-the-art models score so close to the benchmark’s ceiling that their results are virtually indistinguishable. At that point, the test no longer differentiates between new systems, so researchers usually stop reporting scores as the benchmark no longer informs about improvements from the next generation of models. In the OpenAI announcement of GPT-5, they reported results on six flagship public benchmarks (AIME 2025, SWE-bench Verified, Aider Polyglot, MMMU, HealthBench Hard, GPQA) but the full system-card covers roughly thirty-five separate evaluations, comprising hundreds of test task items in total. There have been some efforts to summarize benchmarks in specific fields, like for text-to-image generation, but these surveys have had a narrow methodology scope. Therefore, a comprehensive survey of all benchmarks or even all benchmarks that could be relevant to Steel Thread is outside of the scope of this report. We chose some specific benchmarks to investigate in detail.