Ryan Marten

Ryan Marten is an AI researcher and engineer at the Laude Institute developing the training data, benchmarks, and infrastructure behind more capable reasoning models and autonomous agents. A former founding engineer at Bespoke Labs, he has helped lead OpenThoughts and Terminal-Bench and contributed to Harbor and OpenThoughts-Agent.

Marten studied computer science at the University of Toronto and earned a master’s in computer science from the University of Illinois Urbana-Champaign, where he worked with Derek Hoiem and received a Siebel Scholarship. His research experience includes the Vector Institute, Oxford’s Torr Vision Group, and the Allen Institute for Artificial Intelligence. At AI2, he contributed to DataComp, which examined how dataset choices affect multimodal models, and coauthored Unified-IO 2, which integrated images, language, audio, and actions.

At Bespoke Labs, Marten helped coordinate OpenThoughts, contributing to training, data generation, evaluation, and experimentation. More than 1,000 controlled experiments informed OpenThoughts3, a 1.2-million-example open reasoning dataset, and OpenThinker3, a seven-billion-parameter model evaluated across mathematics, coding, and science.

His most distinctive contributions and technical positions include:

• Data recipes over raw scale. Strong reasoning datasets depend on question sourcing, selective filtering, teacher selection, and repeated evaluation. Synthetic questions can outperform weaker human-generated material; difficulty labels help identify valuable coding problems, while response length provides a stronger signal for mathematics and science.

• Better models are not always better teachers. In OpenThoughts experiments, QwQ-32B produced more effective training data than DeepSeek-R1 despite the latter’s stronger benchmark performance. Sampling multiple reasoning traces per question also increased useful training data without requiring proportionally more distinct problems.

• Supervised fine-tuning can teach reasoning. Training on detailed reasoning traces gives smaller models a practical complement to reinforcement learning. In a legal-reasoning experiment described in his OpenThoughts presentation, a specialized seven-billion-parameter model surpassed its larger teacher after training on verified examples.

• Realistic environments make agent progress measurable. Marten coauthored Terminal-Bench, which evaluates agents on command-line tasks using executable environments and verification tests, and joined the Terminal-Bench 3 project leadership. He has also contributed to Harbor, an open framework for running agents in sandboxed environments.

• OpenThoughts-Agent extends reasoning-data experimentation into agent training. Marten helped coordinate the project, whose research paper reports more than 100 ablation experiments, a 100,000-example training dataset, and 44.8 percent average accuracy across seven agent benchmarks.