/Tech10h ago

Pythagoras-Prover-4B launches, beating the 671-billion parameter DeepSeek-Prover-V2 on the MiniF2F theorem-proving benchmark

A larger 32B variant achieved a new benchmark state-of-the-art.

1063153630K

#714

Original post

Joshua Ong @ ICML@joshuaongg21

🚀🚀 Introducing Pythagoras-Prover 🚀🚀

🔹 The smallest theorem prover to date – Pythagoras-Prover-4B 🔹 The first proof-of-concept diffusion theorem proving model – Pythagoras-Prover-Diffusion 4B

With efficient LoRA training and an 8192-token context length, Pythagoras-Prover achieves strong MiniF2F results with far smaller models.

🏆 MiniF2F - Pythagoras-Prover-4B achieves 86.1% Pass@32, outperforming DeepSeek-Prover-V2-671B. - Pythagoras-Prover-32B achieves 89.8% Pass@32 and 92.6% Pass@2024, achieving the best MiniF2F result reported to date.

🌏Website: https://pythagoras-lm.github.io/ 📝 arXiv: https://arxiv.org/pdf/2606.12594 🤖 Pythagoras-Prover-4B: https://huggingface.co/Pythagoras-LM/Pythagoras-Prover-4B

🔐 Our models, data, and the training pipeline will be released (gradually) soon.

A huge thank to my amazing collaborators: @zhengzhao97 @CMihaelaStoian Qiyuan Xu, @haonanlp @WendaLi8 Shay Cohen @e_giunchiglia

9:24 AM · Jun 15, 2026 · 27.1K Views

Sentiment

Sentiment building, check back later.

Cluster Engagement

Posts from X

Most Activity

VIEWS5.5KBOOKMARKS26LIKES36RETWEETS11REPLIES2

Rohan Paul@rohanpaul_ai

Pythagoras-Prover just made Lean theorem proving look far less dependent on giant models, with a 4B prover beating DeepSeek-Prover-V2-671B at MiniF2F Pass@32.

Shows in formal reasoning, better data geometry can buy back an astonishing amount of scale.

A theorem prover is not just a language model writing clever math; it is a machine trying to produce text that survives a compiler with no patience for style, confidence, or almost-right reasoning.

The main trick is data efficiency: the team built about 800K Lean-verified examples, trained from easy to hard, then used LoRA so the model learned without updating every parameter.

Joshua Ong @ ICML@joshuaongg21

🚀🚀 Introducing Pythagoras-Prover 🚀🚀

🔹 The smallest theorem prover to date – Pythagoras-Prover-4B 🔹 The first proof-of-concept diffusion theorem proving model – Pythagoras-Prover-Diffusion 4B

With efficient LoRA training and an 8192-token context length, Pythagoras-Prover achieves strong MiniF2F results with far smaller models.

🌏Website: https://pythagoras-lm.github.io/ 📝 arXiv: https://arxiv.org/pdf/2606.12594 🤖 Pythagoras-Prover-4B: https://huggingface.co/Pythagoras-LM/Pythagoras-Prover-4B

🔐 Our models, data, and the training pipeline will be released (gradually) soon.

A huge thank to my amazing collaborators: @zhengzhao97 @CMihaelaStoian Qiyuan Xu, @haonanlp @WendaLi8 Shay Cohen @e_giunchiglia

3h5.5K3626

Rohan Paul@rohanpaul_ai

https://arxiv.org/pdf/2606.12594

Rohan Paul@rohanpaul_ai

Pythagoras-Prover just made Lean theorem proving look far less dependent on giant models, with a 4B prover beating DeepSeek-Prover-V2-671B at MiniF2F Pass@32.

Shows in formal reasoning, better data geometry can buy back an astonishing amount of scale.

A theorem prover is not just a language model writing clever math; it is a machine trying to produce text that survives a compiler with no patience for style, confidence, or almost-right reasoning.

The main trick is data efficiency: the team built about 800K Lean-verified examples, trained from easy to hard, then used LoRA so the model learned without updating every parameter.

3h2.3K50

Joshua Ong @ ICML@joshuaongg21

(1/5) 🧩 At the core is a compute-frugal synthetic data pipeline: we build a Lean-verified corpus stratified into easy, medium, and hard tiers, then train with a difficulty curriculum — letting small models learn proof skills progressively, from short proofs to hard olympiad problems.

10h1251

Joshua Ong @ ICML@joshuaongg21

(2/5) ✨ We introduce Augmented Lean Formalisation (ALF): a way to augment Lean formalisations at scale without per-instance verification — removing the dominant cost of formal data generation and helping solve the scarcity of open-source Lean data. Reused via self-distillation, it pushes the performance of both our 4B and 32B provers.

10h881

Joshua Ong @ ICML@joshuaongg21

(4/5) ⚡ Finally, a proof of concept for diffusion-based proving. We identify what holds diffusion provers back (sparse masking signal, limited stable context, scarce diffusion-ready data), mitigate them with tactic-based masking + ALF, and show it generates proofs 2.58× faster than its AR counterpart.

10h831

Joshua Ong @ ICML@joshuaongg21

(3/5) 🔍 We run a thorough analysis of how current provers behave on MiniF2F — and find it's saturating. So we release MiniF2F-ALF, a perturbed variant that reopens the benchmark, diversifies the failure modes, and better separates strong systems.

10h511

Joshua Ong @ ICML@joshuaongg21

(5/5) 📄 Full details are on our website and on arXiv (https://pythagoras-lm.github.io/). I'll also be at ICML 2026🇰🇷 this year — reach out if you'd like to chat about Pythagoras-Prover, formal reasoning, or diffusion for proving! 🤝

10h731