Research · Research Lead

Carlo Baronio

“Research @ Cognition”

LinkedIn X

Joined: Apr 2025
How they joined: Directly
Before Cognition: Stanford Artificial Intelligence Laboratory (SAIL)Undergraduate Researcher @ Hazy Research
Olympiads: IMO bronze 2023IMO 2023Bronze

Overlapped before Cognition

Stanford University16

Sources¹linkedin.com

Autumn profile

Researched with Autumn · 2026-11-04

Details

LocationStanford, California, United States

Company sitecognition.ai

UniversityStanford University

Notes

Promoted to Research Lead at Cognition in October 2026 after ~18 months as a Research Engineer (from April 2025).Oct 2026
Co-author of SWE-2 (Sept 2026); posted verbatim: "SWE-2 is our first attempt to scale intelligence/cost and on FrontierCode it matches Fable 5.1 at 64% lower cost. We achieved this by developing a penalty that preserves the shape of the test-time scaling curve during training.".Sep 10, 2026SWE-2 is our first attempt to scale intelligence/cost and on FrontierCode it matches Fable 5.1 at 64% lower cost. We achieved this by developing a penalty that preserves the shape of the test-time scaling curve during training.
Co-lead (equal-contribution author, listed second after Ben Pan) of SWE-1.7, Cognition's July 2026 model shipped in Windsurf that the blog says 'reaches frontier-level intelligence at a much lower cost'.Jul 8, 2026
First author (equal contribution with Pietro Marsella and Ben Pan, advised by Silas Alberti) of Kevin-32B, the first open model trained with multi-turn RL to write CUDA kernels, built as a Stanford-Cognition collaboration.May 6, 2025
Posted verbatim on SWE-1.7 launch: "It's a good model sir?!! Pls check out our blog post there are a lot of cool findings/stuff we do to get huge gains from just RL on high quality data at scale. There is no wall 🧱"Jul 8, 2026It's a good model sir?!! Pls check out our blog post there are a lot of cool findings/stuff we do to get huge gains from just RL on high quality data at scale. There is no wall 🧱 https://t.co/jHR0MXrnFU
Co-author of SWE-1.6 (April 2026), released generally available in Windsurf, the IDE Cognition acquired in July 2025; his model work therefore ships directly into the acquired Windsurf product (Devin Desktop/Windsurf line).Apr 7, 2026
Co-author of SWE-grep and SWE-grep-mini (Oct 2025), fast RL-trained multi-turn context-retrieval models powering Windsurf's Fast Context subagent.Oct 16, 2025
Current role is Research Lead at Cognition.
Co-author of FrontierCode (June 2026), Cognition's hard coding eval, and listed in Research acknowledgments of FrontierCode 1.1 (July 2026), which became the benchmark SWE-2 claims are measured on.Jun 8, 2026
Posted verbatim on SWE-1.6 preview: "Nice milestone for our research effort: SWE-1.6 (early preview) is a meaningful step up from 1.5 without sacrificing speed (950 tok/s). Strong SWE-Bench Pro results, and this is just the beginning!"Mar 1, 2026Nice milestone for our research effort: SWE-1.6 (early preview) is a meaningful step up from 1.5 without sacrificing speed (950 tok/s). Strong SWE-Bench Pro results, and this is just the beginning!
Kevin paper (with Simon Guo of Stanford/KernelBench added as co-author) was accepted at ICLR 2026; Google Scholar shows ~117 citations for it.2026
Featured speaker at CoreWeave Fully Connected 2026, presenting on training Cognition's SWE models with asynchronous RL at global scale; CoreWeave's bio says he works on algorithms and training infrastructure spanning Kevin-32B, SWE-grep and Cognition's SWE models.2026

Autumn profile

People they worked with before

More from Research