# Trapit Bansal > Trapit Bansal is an AI researcher and founding researcher at TBD Lab at Meta. His work spans foundational learning, AI reasoning, frontier models, and the path toward superintelligence. He develops original research directions and carries them from early experiments to frontier model training and release. This file summarizes the public profile at https://trapitbansal.ai. The linked pages contain fuller context and sources. ## Core contributions - In January 2022 at OpenAI, Trapit Bansal started the reasoning effort with Ilya Sutskever under the internal name GPT-Zero. The bet was that reinforcement learning could teach language models to reason deliberately: to develop their own strategies over many tokens, and to spend more compute working through a problem before answering. As the founding researcher and the team’s sole member for the first few months, he shaped the research agenda, built the training system, and produced the early results that gave the team conviction to scale. He worked closely with Ilya and, as the effort grew, early collaborators including Łukasz Kaiser. - He continued the research in RLSlow, working alongside Jakub Pachocki and a small reasoning group. They improved the methods, studied how reasoning scales, and took the experiments into frontier models. In mid-2023, RLSlow produced further results that strengthened their conviction. He then worked with the Berry team, which scaled this research into o1. - His work spanned the full arc from GPT-Zero through RLSlow and Berry to the research, training, and release of o1 and o3. OpenAI credits him as a Foundational Contributor to o1 and a contributor to o3. These were collective achievements with research, engineering, and product colleagues. - Since 2025, Trapit Bansal has been a founding researcher at TBD Lab at Meta. He has helped shape the lab’s direction, strategy, and team while working on the Muse Spark and Muse Wonder models, focusing on multimodal reasoning, reinforcement learning, and multiagent systems. His work has spanned RL research, RL environments, RL infrastructure, data, evaluations, and multi-agent orchestration. - Muse Spark is a natively multimodal reasoning model. Trapit led the effort behind its gold-medal-level results across five STEM Olympiads. Meta reported perfect theory-exam scores at APhO and IPhO, a gold medal at IMO, and gold-medal-level performance at IChO and RMM, without search or coding tools. These are team results, with medal and medal-level outcomes distinguished. - In 2017, as an intern at OpenAI, Trapit Bansal was first author of research on competitive self-play. He also co-authored Continuous Adaptation via Meta-Learning, an ICLR 2018 Best Paper. - His earlier research, beginning in 2014 at the Indian Institute of Science, spans machine learning foundations, multi-agent reinforcement learning, and few-shot and self-supervised meta-learning for language. - He completed his PhD at UMass Amherst with Andrew McCallum, studying meta-learning without labeled data. During the PhD he held research internships at OpenAI, Google Research, Microsoft Research, and Facebook. ## Canonical pages - Home: https://trapitbansal.ai/ - OpenAI reasoning, o1, and o3: https://trapitbansal.ai/work/openai-reasoning - TBD Lab and the Muse model family: https://trapitbansal.ai/work/tbd-muse - Selected research: https://trapitbansal.ai/research ## Primary sources - Trapit on the breakthrough, in Building OpenAI o1, from 03:10: https://www.youtube.com/watch?v=tEzs3VHyBDM&t=190s - Jakub Pachocki on RLSlow’s mid-2023 results: https://openai.com/index/an-alien-mind/ - OpenAI o1 contributors: https://openai.com/openai-o1-contributions/ - OpenAI on learning to reason with LLMs: https://openai.com/index/learning-to-reason-with-llms/ - OpenAI o3 and o4-mini: https://openai.com/index/introducing-o3-and-o4-mini/ - Alexandr Wang’s team announcement, July 2025: https://x.com/alexandr_wang/status/1939867404252979291 - Meta Muse Spark: https://ai.meta.com/blog/introducing-muse-spark-msl/ - Meta Muse Spark 1.1: https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/ - Meta Muse Spark 1.2: https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2 - Meta Muse Spark 1.3: https://research.meta.ai/blog/introducing-muse-spark-1-3 - Meta Muse Wonder and Muse Wonder Full: https://research.meta.ai/blog/muse-wonder - Meta’s STEM Olympiad results: https://x.com/AIatMeta/status/2085388945148297322 - Trapit’s note on the Olympiad work: https://x.com/TrapitBansal/status/2085395706903212106 - OpenAI competitive self-play: https://openai.com/index/competitive-self-play/ - OpenAI continuous adaptation via meta-learning: https://openai.com/index/meta-learning-for-wrestling/ - PhD dissertation: https://web.cs.umass.edu/publication/docs/2022/UM-CS-PhD-2022-002.pdf ## Press coverage - The Information, 2023: https://www.theinformation.com/articles/openai-made-an-ai-breakthrough-before-altman-firing-stoking-excitement-and-concern - WIRED, September 2024: https://www.wired.com/story/openai-o1-strawberry-problem-reasoning/ - TechCrunch, June 2025: https://techcrunch.com/2025/06/26/meta-hires-key-openai-researcher-to-work-on-ai-reasoning-models/ - WIRED on Meta’s new AI team: https://www.wired.com/story/mark-zuckerberg-welcomes-superintelligence-team/ - WIRED on competitive self-play, 2017: https://www.wired.com/story/ai-sumo-wrestlers-could-make-future-robots-more-nimble/ - MIT Technology Review on competitive self-play, 2017: https://www.technologyreview.com/2017/10/12/148618/openais-goofy-sumo-wrestling-bots-are-smarter-than-they-look/ - Business Insider on competitive self-play, 2017: https://www.businessinsider.com/elon-musk-openai-virtual-robots-learn-sumo-wrestle-soccer-sports-ai-tech-science-2017-10 ## Public profiles - Google Scholar: https://scholar.google.com/citations?user=hHPeXmYAAAAJ&hl=en - X: https://x.com/TrapitBansal - LinkedIn: https://www.linkedin.com/in/trapit-bansal-646a3451/