- Joined
- Apr 2025
- How they joined
- Directly
- Before Cognition
- Gordian SoftwareSoftware Engineer
- Status
- Current
Overlapped before Cognition
“One year ago I joined Cognition to train the best coding model. We were a handful of undergrad and PhD dropouts, and part of me wondered whether we could really get to the frontier without the experience.”
Sources1linkedin.com
Details
reinforcement learningLLM post-trainingtraining data researchCUDA kernel generationcompetitive programming
Notes
- One-year anniversary post (verbatim): "One year ago I joined Cognition to train the best coding model. We were a handful of undergrad and PhD dropouts, and part of me wondered whether we could really get to the frontier without the experience." He adds the team is now "20+ world-class research engineers".Sep 10, 2026One year ago I joined Cognition to train the best coding model. We were a handful of undergrad and PhD dropouts, and part of me wondered whether we could really get to the frontier without the experience. Today we’re a team of 20+ world-class research engineers, and we just ship
- Listed first (with asterisk) on SWE-1.7, 'Frontier Intelligence at a Fraction of the Cost' (Ben Pan*, Carlo Baronio*, Rohan Choudhury, Eric Lu, Ryan Kim and others).Jul 8, 2026
- His launch post for SWE-1.7 credits data research: "It's been especially rewarding to watch our data research grow from early experiments to materialize in concrete model behaviors"; colleague Cheng-Yuan (Sam) Lee replied "ben is the best data researcher".Jul 8, 2026SWE-1.7 is the first frontier-level model from our research team. Over the last three months, we scaled our RL across four axes: training steps, task horizon, compute, and data.
- Co-authored Kevin-32B (K(ernel D)evin), Cognition's first multi-turn RL model for writing CUDA kernels, with Carlo Baronio, Pietro Marsella and Silas Alberti; the blog marks the authors with equal-contribution asterisks.May 6, 2025
- Joined Cognition as a Research Engineer in April 2025 while still a Stanford math/CS undergrad (Stanford 2023-2027 on LinkedIn).Apr 2025
- SWE-2 (Sep 2026) scaled RL to the multi-trillion-parameter regime 'building on the SWE-1.7 training infrastructure and recipe', so the pipeline he co-led became the base of Cognition's next flagship model.Sep 10, 2026
- Same post: "Turns out there's no secret sauce. A small team of very smart people can figure out anything from first principles." The SWE-2 post drew congratulations from Sarah Guo (Conviction), Alex Stauffer and nathan chen, a visible investor-adjacent audience.Sep 10, 2026One year ago I joined Cognition to train the best coding model. We were a handful of undergrad and PhD dropouts, and part of me wondered whether we could really get to the frontier without the experience. Today we’re a team of 20+ world-class research engineers, and we just ship
- Co-author on FrontierCode, Cognition's coding benchmark (Eric Lu, Ben Pan, Deniz Birlikci, Sam Lee, Ray Wang, Rohan Choudhury, Fermi Ma, TC Qin, Carlo Baronio, Silas Alberti), cited in the RECLAIM paper; FrontierCode v1.1 was later adopted in Artificial Analysis's Coding Agent Index v2.0.2026
- Co-authored SWE-grep and SWE-grep-mini, Cognition's RL-trained multi-turn fast context-retrieval models for Windsurf, with Carlo Baronio, Albert Tam, Pietro Marsella and others.Oct 2025
- Cognition signed a definitive agreement to acquire Windsurf (agentic IDE, $82M ARR, 350+ enterprise customers) on 2025-07-14, three months after he joined; his later SWE-1.6/1.7 models ship inside Windsurf and Devin, so the acquisition gave his research team a distribution surface.Jul 14, 2025
- The Kevin paper (arXiv 2507.11948) lists him with dual affiliation Stanford University and Cognition AI, alongside Carlo Baronio, Pietro Marsella, Simon Guo and Silas Alberti; it reports correctness rising from 56% to 82% over base QwQ-32B and mean speedup from 0.53x to 1.10x.Jul 2025
- Gary (Bin) Chang, who came to Cognition via the Windsurf acquisition, named Ben Pan among his current research teammates with Silas Alberti, Cheng-Yuan Lee, Ray Wang, Carlo Baronio, Connor Fogarty and Emile Timothy Anand, showing ex-Windsurf engineers were folded into the same model-training team.Jan 2026
Experience
Research Engineer at Cognition
Apr 2025 - Present
Software Engineer at Gordian Software
Jul 2021 - Aug 2021
YC W19
Education
Stanford University
Jan 2023 - Jan 2027
Mathematics and Computer Science
Projects
AchievementsCo-led SWE-1.7, Cognition's first frontier-level model · Kevin paper accepted to ICLR 2026 · 2nd place, PClassic advanced division (UPenn) with Stuyvesant team · Cooke College Scholarship semi-finalist (2023)
Recent posts
July 2026
SWE-1.7 is the first frontier-level model from our research team. Over the last three months, we scaled our RL across four axes: training steps, task horizon, compute, and data.
September 2026
One year ago I joined Cognition to train the best coding model. We were a handful of undergrad and PhD dropouts, and part of me wondered whether we could really get to the frontier without the experience. Today we’re a team of 20+ world-class research engineers, and we just ship
Historical
Articles
AwardsCooke College Scholarship semi-finalist · PClassic (UPenn) advanced division 2nd place
Profiles
