Yuhuai Wu

Yuhuai (Tony) Wu is a Chinese-Canadian machine-learning researcher known for work on neural theorem proving, autoformalization, and machine reasoning systems for mathematics. He is a co-author of "Solving Olympiad Geometry Without Human Demonstrations" (the AlphaGeometry paper, Nature, January 2024), the Minerva quantitative-reasoning system, the Memorizing Transformer, and the Autoformalization with Large Language Models paper, and he was one of the eleven publicly named founding team members of xAI from March 2023 until his departure on February 10, 2026. As of May 2026, his post-xAI plans have not been publicly disclosed.

At a glance

• Education: BSc in Mathematics, University of New Brunswick, 2015. PhD in the Department of Computer Science at the University of Toronto, thesis "Neural Networks for Mathematical Reasoning" deposited March 2024, advised by Roger Grosse (advisor) and Jimmy Ba (co-advisor). Postdoctoral researcher at Stanford with Percy Liang and James McClelland.
• Previous role: Founding team member of xAI, March 2023 to February 10, 2026. Press coverage credited him with leadership on the Grok 3 reasoning effort.
• Pre-xAI industry: Research Scientist at Google Research, on the N2Formal team led by Christian Szegedy, 2021 to 2023.
• Key contributions: AlphaGeometry (Trinh, Wu, Le, He, Luong, Nature 2024); Minerva (Lewkowycz et al., NeurIPS 2022); Autoformalization with LLMs (Wu et al., NeurIPS 2022); Memorizing Transformers (Wu, Rabe, Hutchins, Szegedy, ICLR 2022); Draft, Sketch, and Prove (Jiang, Welleck, Zhou, Liu, Li, Lin, Wu, Sutskever, et al., ICLR 2023).
• Awards: Google PhD Fellowship (2017 to 2020); NSERC Canada Graduate Scholarship (Doctoral); University of New Brunswick BSc with first-class honors (2015).
• X / Twitter: @Yuhu_ai_
• Personal site: yuhuaiwu.github.io
• GitHub: tonywu95
• Google Scholar: Yuhuai (Tony) Wu (h-index 42, more than 35,000 citations as of May 2026)

Origins

Public biographical material on Wu is comparatively thin. He has no Wikipedia entry as of May 2026, and the available record runs through his personal site at yuhuaiwu.github.io, his X account, his Google Scholar profile, his University of Toronto faculty pages, the TSpace deposit of his PhD thesis, the published papers, and press coverage of the February 2026 xAI departure. He goes publicly by both "Yuhuai Wu" and "Tony Wu"; the parenthetical "(Tony)" appears on his personal site, his published papers, and his GitHub identity.

Wu was born in 1995 in Jiande, Zhejiang Province, China. He attended primary school in Zhejiang and then Hangzhou before relocating to Canada for high school. His undergraduate years at the University of New Brunswick were in mathematics; he graduated with a BSc in 2015. The transition into machine learning followed an unsolicited 2014 email he sent as a pure-mathematics student to Yoshua Bengio, who responded by offering him a summer internship at Mila in 2015. Wu has cited that internship as the start of his ML career.

Career

Wu joined the University of Toronto Department of Computer Science as a PhD student in 2015, under the joint supervision of Roger Grosse and Jimmy Ba. The thesis acknowledgements identify Christian Szegedy as his mentor in the field of mathematical reasoning, dating the relationship to a Google internship; Szegedy and Wu would later overlap at xAI. A summer 2018 visit to DeepMind put Wu in contact with James McClelland, whose talk on mathematical reasoning convinced Wu to focus the thesis on the topic and who later helped recruit him to a Stanford postdoc with Percy Liang. The thesis "Neural Networks for Mathematical Reasoning: Evaluations, Capabilities, and Techniques" was deposited at TSpace in March 2024, with the degree formally awarded in 2024.

The PhD period produced the neural-theorem-proving line that has run through Wu's subsequent career. "INT: An Inequality Benchmark for Evaluating Generalization in Theorem Proving" (Wu, Jiang, Ba, Grosse, ICLR 2021) introduced an inequality benchmark for generalization in neural theorem proving. "LIME: Learning Inductive Bias for Primitives of Mathematical Reasoning" (Wu, Rabe, Li, Ba, Grosse, Szegedy, ICML 2021) examined synthetic pretraining tasks that injected useful inductive bias into transformer models targeting mathematical reasoning.

Wu joined Google Research as a Research Scientist in 2021 on the N2Formal team led by Christian Szegedy. "Memorizing Transformers" (Wu, Rabe, Hutchins, Szegedy, ICLR 2022 Spotlight) introduced an external memory architecture extending transformer context to handle large mathematical libraries. "Autoformalization with Large Language Models" (Wu, Jiang, Li, Rabe, Staats, Jamnik, Szegedy, NeurIPS 2022) reported that LLMs could correctly translate roughly 25 percent of competition-math problems into machine-checkable Isabelle/HOL specifications without supervised training, framing autoformalization as a viable program for AI-driven formal mathematics. "Draft, Sketch, and Prove" (Jiang et al., ICLR 2023 Oral) extended the program by training models to produce informal proof sketches that guide formal-prover search. Wu was also a co-author on "Solving Quantitative Reasoning Problems with Language Models" (Lewkowycz et al., NeurIPS 2022), introducing Minerva, the language model trained for technical and quantitative reasoning, and on the PaLM 2 technical report (Anil et al., 2023).

In March 2023, Wu left Google to join xAI as one of the eleven publicly named founding team members. The team was assembled from senior researchers at Google DeepMind, OpenAI, and Google Brain, and the launch was publicly announced on July 12, 2023. Press coverage of his subsequent departure characterized Wu as a leader on the Grok 3 reasoning effort. Specific functional titles beyond "founding team member" are not publicly stated.

In January 2024, "Solving Olympiad Geometry Without Human Demonstrations" (Trinh, Wu, Le, He, Luong, Nature 2024) reported that the AlphaGeometry system solved 25 of 30 International Mathematical Olympiad geometry problems within standard contest time limits, a gold-medalist-level performance. The work was carried out at Google DeepMind, with Wu listed among the principal authors.

On February 10, 2026, Wu posted on X that he had resigned from xAI. The post described the company and founding team as a "family", thanked colleagues for shared work, and framed the next chapter as one in which "a small team armed with AIs can move mountains and redefine what's possible." Bloomberg, Silicon Republic, and TechCrunch covered the announcement, characterizing the departure as the sixth among the original twelve named co-founders. Jimmy Ba announced his own departure roughly seventeen hours later. The departures came shortly after the SpaceX acquisition of xAI on February 2, 2026, in an all-stock transaction valuing the combined entity at $1.25 trillion. Press coverage cited tensions between xAI's research orientation and the operational intensity following the SpaceX integration. As of May 2026, Wu's post-xAI plans have not been publicly disclosed beyond the framing in his X post.

Affiliations

• University of New Brunswick: BSc in Mathematics, 2011 to 2015.
• Mila: Research intern under Yoshua Bengio, summer 2015.
• University of Toronto, Department of Computer Science: PhD candidate, 2015 to 2024 (degree formally awarded 2024), advised by Roger Grosse and co-advised by Jimmy Ba.
• Google DeepMind: Research intern, summer 2018.
• Stanford University: Postdoctoral researcher with Percy Liang and James McClelland, approximately 2021.
• Google Research: Research Scientist, N2Formal team, 2021 to 2023.
• xAI: Founding team member, March 2023 to February 10, 2026.

Notable contributions

Wu's published record concentrates in neural theorem proving, autoformalization, mathematical-reasoning benchmarks, and large-language-model methods for formal and informal mathematics. His Google Scholar profile lists the foundation-models survey, AlphaStar, HELM, PaLM 2, and Minerva among the most-cited entries he has co-authored, with the autoformalization and theorem-proving papers as his principal first- and joint-first-authored contributions.

• "Solving Olympiad Geometry Without Human Demonstrations" (Trinh, Wu, Le, He, Luong, Nature, January 2024). Co-author paper introducing AlphaGeometry, the first AI system to reach IMO-gold-medalist performance on a benchmark of olympiad geometry problems.
• "Autoformalization with Large Language Models" (Wu, Jiang, Li, Rabe, Staats, Jamnik, Szegedy, NeurIPS 2022). Co-first-author paper showing that LLMs can directly translate competition mathematics into Isabelle/HOL formal specifications.
• "Memorizing Transformers" (Wu, Rabe, Hutchins, Szegedy, ICLR 2022 Spotlight). First-author paper on extending transformer context with non-differentiable memory, with applications to formal-mathematics corpora.
• "Draft, Sketch, and Prove: Guiding Formal Theorem Provers with Informal Proofs" (Jiang et al., ICLR 2023 Oral). Co-author paper on generating informal proof sketches to guide formal proof search.
• "Solving Quantitative Reasoning Problems with Language Models" (Lewkowycz, Andreassen, Dohan, Wu, et al., NeurIPS 2022). Co-author paper introducing Minerva, the language model trained for technical and quantitative reasoning.
• "STaR: Bootstrapping Reasoning With Reasoning" (Zelikman, Wu, Mu, Goodman, NeurIPS 2022). Co-author paper introducing self-taught reasoner training, in which a language model generates and filters its own reasoning traces.
• "INT: An Inequality Benchmark for Evaluating Generalization in Theorem Proving" (Wu, Jiang, Ba, Grosse, ICLR 2021). First-author paper introducing the INT benchmark for inequality theorem proving.
• "LIME: Learning Inductive Bias for Primitives of Mathematical Reasoning" (Wu, Rabe, Li, Ba, Grosse, Szegedy, ICML 2021). First-author paper on synthetic pretraining tasks for mathematical-reasoning transformer models.
• xAI founding-team research contributions (March 2023 to February 2026). Press coverage credited Wu with leadership on the Grok 3 reasoning effort. Specific authored xAI artifacts are not publicly disclosed.

Investments and boards

No public investor activity on record in AI, semiconductors, datacenters, software, or energy as of May 2026.

Network

Wu's longest-running professional relationships fall in three cohorts. The first is the University of Toronto machine-learning group, including his thesis advisors Roger Grosse and Jimmy Ba and frequent co-author Albert Q. Jiang on the formal-mathematics line. The second is the Google Research N2Formal team and the formal-mathematics community, including Christian Szegedy (Wu's mentor in the field, later xAI co-founder, now founder of Math Inc), Markus Rabe, Wenda Li, DeLesley Hutchins, and the AlphaGeometry team at Google DeepMind led by Trieu Trinh and Thang Luong. The third is the xAI founding team. Beyond Elon Musk, the founding cohort included Igor Babuschkin (departed August 2025 to launch Babuschkin Ventures), Greg Yang (informal advisor since January 2026), Christian Szegedy (departed February 2025 to join Morph Labs and later found Math Inc), Jimmy Ba (departed February 10, 2026), Manuel Kroiss, Toby Pohlen, Ross Nordeen, Kyle Kosic, Guodong Zhang, and Zihang Dai. Yoshua Bengio remains a long-standing mentor since the 2015 Mila internship.

Position in the field

As of May 2026, Wu occupies a structurally distinctive position in the mathematical-reasoning research line. The combination of co-authorship on AlphaGeometry, on Minerva, on the autoformalization paper, and on the Memorizing Transformer places him among the small group of researchers most associated with the contemporary push to make large language models capable of formal and informal mathematical reasoning. The Google Scholar citation count above 35,000 is high for the post-PhD career stage. Press coverage of his February 2026 departure from xAI consistently led with the AlphaGeometry and autoformalization attribution.

His work is the most direct intellectual lineage between the academic neural-theorem-proving line and the production reasoning capabilities now embedded in frontier large language models. The autoformalization research program he helped frame at Google has continued at xAI through the Grok reasoning effort, at Math Inc under Christian Szegedy, and at Google DeepMind through AlphaProof and AlphaGeometry 2. His public-commentary cadence is moderate: he posts intermittently to @Yuhu_ai_, with limited solo English-language video presence beyond the NeurIPS Chats episode on Minerva and bootstrapping reasoning, the December 2023 DAI keynote on autoformalization in Singapore, and academic seminar appearances.

Outlook

Open questions over the next 6 to 18 months:

• Post-xAI role. Whether Wu starts a new venture aligned with the framing in the February 10 X post, joins another industry research lab, or returns to academic publication cadence. The departure post pointed toward a small-team-and-AIs configuration without naming a destination.
• xAI relationship. Whether the post-departure relationship remains technically active, given the Grok 3 reasoning lineage and the prior collaboration with Christian Szegedy at Google.
• Autoformalization trajectory. Whether Wu continues to publish on the autoformalization line that links his Google, xAI, and the broader formal-mathematics community, including the Lean and Mathlib ecosystems.
• Math Inc collaboration. Whether the working relationship with Szegedy at Math Inc resumes commercially or publicly, given the shared autoformalization research program.
• Public-commentary cadence. Whether Wu's posting cadence and his seminar-and-podcast schedule pick up after the xAI departure.

Sources

• Yuhuai (Tony) Wu (personal site). Personal site listing his research focus, publications, and contact information.
• Yuhuai (Tony) Wu on X. The @Yuhu_ai_ X account, the primary public channel for his commentary and the source of the February 10, 2026 resignation announcement.
• Yuhuai (Tony) Wu on Google Scholar. Publication and citation listing.
• Neural Networks for Mathematical Reasoning. The University of Toronto PhD thesis by Wu, deposited March 2024 and advised by Roger Grosse and co-advised by Jimmy Ba.
• tonywu95 on GitHub. Public GitHub profile.
• Yuhuai Wu's resignation post on X. The February 10, 2026 announcement of his departure from xAI.
• Tony Wu Becomes Latest xAI Co-Founder to Leave Musk's Startup. Bloomberg coverage of the February 10, 2026 departure.
• Jimmy Ba and Tony Wu latest tech co-founders to exit Musk's xAI. Silicon Republic coverage of the founding-team exits and the SpaceX-acquisition timing.
• Okay, now exactly half of xAI's founding team has left the company. TechCrunch coverage of the Wu and Ba departures.
• Musk's xAI loses second co-founder in two days as Jimmy Ba departs. CNBC coverage situating the Wu and Ba departures.
• Solving Olympiad Geometry Without Human Demonstrations. The January 2024 Nature paper introducing AlphaGeometry, with Wu as a principal co-author.
• AlphaGeometry: An Olympiad-level AI system for geometry. Google DeepMind blog post on the AlphaGeometry result.
• Autoformalization with Large Language Models. The 2022 NeurIPS paper introducing autoformalization with large language models.
• Memorizing Transformers. The 2022 ICLR Spotlight paper on transformer external memory.
• Draft, Sketch, and Prove: Guiding Formal Theorem Provers with Informal Proofs. The 2023 ICLR Oral paper on informal-proof-guided formal theorem proving.
• Solving Quantitative Reasoning Problems with Language Models. The 2022 NeurIPS paper introducing Minerva.
• STaR: Bootstrapping Reasoning With Reasoning. The 2022 NeurIPS paper on self-taught reasoning.
• INT: An Inequality Benchmark for Evaluating Generalization in Theorem Proving. The 2021 ICLR paper introducing the INT benchmark.
• LIME: Learning Inductive Bias for Primitives of Mathematical Reasoning. The 2021 ICML paper on synthetic mathematical-reasoning pretraining tasks.
• Announcing xAI. The July 12, 2023 founding announcement, naming Wu among the founding team members.
• NeurIPS Chats: Yuhuai Wu on LMs for Math (Minerva) + the Future of Bootstrapping Reasoning. The NeurIPS Chats interview on Minerva and the bootstrapping-reasoning research line.
• Autoformalization with Large Language Models (DAI 2023 keynote). The December 2023 DAI keynote listing on autoformalization.