Algorithmic Research Group

Algorithmic Research Group (ARG) is a small AI safety research lab founded in October 2024 by Matthew Kenney and headquartered in Durham, North Carolina. The organization's central research focus is understanding recursive self-improvement: how software and industrial systems iteratively optimize themselves in real-world settings, and what the behavior and limits of such systems look like in practice. ARG builds benchmarks, datasets, and agentic infrastructure to study these dynamics empirically. Their most notable publication is the ML Research Benchmark (MLRB), introduced in an October 2024 arXiv paper by Matthew Kenney. The MLRB comprises seven competition-level tasks derived from recent machine learning conference tracks, covering model training efficiency, pretraining on limited data, domain-specific fine-tuning, and model compression. Evaluating Claude-3.5 Sonnet and GPT-4o on these tasks, the paper found that while current AI agents can produce baseline results, they fall short of the capabilities required for advanced AI research — providing a concrete measure of the current frontier. Beyond benchmarking, ARG develops multi-agent environments designed to examine emergent behaviors, GPU kernel optimization tools, neural architecture search systems, and evaluation frameworks for frontier models. They maintain 17 repositories on GitHub and over 30 models and 42 datasets on HuggingFace, with an emphasis on open-source accessibility for the broader research community. ARG also provides commercial services: AI benchmarking, research infrastructure, agentic model design, and consulting for organizations building agentic systems. Their ScoutML product serves as a research agent platform. The organization is very early-stage, with a team of fewer than 10 people and no publicly disclosed external funding.