DeepSeek AI has officially released DeepSeek-Prover-V2, an advanced open-source large language model (LLM) designed for formal theorem proving in the Lean 4 environment. This second-generation model introduces a novel recursive proof-search framework and is trained using synthetic data generated by DeepSeek-V3. It delivers state-of-the-art results on the MiniF2F benchmark and debuts alongside ProverBench, a new evaluation suite for mathematical reasoning.
Cold-Start Training with Self-Generated Data
At the core of DeepSeek-Prover-V2’s innovation is a cold-start training method. The process begins with DeepSeek-V3 breaking down complex theorems into simpler subgoals, which are then formalized into Lean 4 code. These structured proof steps are used to train a smaller 7B parameter model tasked with solving each subgoal independently.
When a full proof is completed through this recursive decomposition, it is paired with DeepSeek-V3’s corresponding chain-of-thought reasoning, resulting in a rich dataset that combines informal insights with formalized logic. This dataset serves as the foundation for bootstrapping the model’s capabilities prior to reinforcement learning.
Reinforcement Learning to Bridge Intuition and Formal Logic
Using the cold-start data, DeepSeek selected challenging theorems where subgoals were provable but the complete solution remained elusive. By assembling the verified subgoal proofs and connecting them to DeepSeek-V3’s reasoning path, the team created holistic training examples that map informal lemma decomposition to formal proofs.
The model was then fine-tuned on this dataset, followed by a reinforcement learning phase. During this stage, the model received binary feedback—correct or incorrect—to refine its understanding and improve its ability to align intuitive reasoning with formal proof construction.
Record-Breaking Results in Theorem Proving
The final product, DeepSeek-Prover-V2–671B, features an impressive 671 billion parameters. It has achieved outstanding performance, including:
- 88.9% pass rate on the MiniF2F-test
- Solved 49 out of 658 problems from PutnamBench
All proofs generated for the MiniF2F dataset are publicly available, enabling the community to explore, validate, and build upon DeepSeek’s results.
DeepSeek-Prover-V2 represents a major leap forward in the intersection of language models and formal mathematics, combining automated reasoning, data synthesis, and reinforcement learning into one cohesive system.
Introducing ProverBench: Setting a New Bar for Mathematical Evaluation
Alongside the release of its latest models, DeepSeek AI has unveiled ProverBench, a newly developed benchmark dataset consisting of 325 formalized problems. This dataset is crafted to provide a more well-rounded and rigorous evaluation of mathematical reasoning across varying levels of complexity.
Among these, 15 problems are derived from recent AIME (American Invitational Mathematics Examination) contests (AIME 24 and 25), reflecting the kind of high-school-level competition challenges that test real-world reasoning under pressure. The remaining 310 problems are carefully selected from textbooks and educational tutorials, ensuring a broad, well-structured mix of topics that span key domains in mathematics.
ProverBench is designed to support in-depth assessment of neural theorem provers, balancing both the difficulty of competition-style questions and the foundational concepts found in undergraduate mathematics.
Model Availability
DeepSeek AI is releasing DeepSeek-Prover-V2 in two sizes to accommodate diverse computational needs:
- 7B parameter model – built on DeepSeek-Prover-V1.5-Base, with extended support for sequences up to 32,000 tokens, making it well-suited for processing longer and more complex proofs.
- 671B parameter model – based on the advanced DeepSeek-V3-Base, offering significantly greater capacity and power.
Together, the launch of DeepSeek-Prover-V2 and the debut of ProverBench represent a major advancement in the realm of neural theorem proving. By integrating a recursive proof search framework and a robust benchmark suite, DeepSeek AI is equipping researchers and developers with the tools needed to push the boundaries of AI in formal mathematics.
Frequently Asked Questions
What is DeepSeek-Prover-V2?
DeepSeek-Prover-V2 is an advanced AI model developed by DeepSeek AI, designed specifically for neural theorem proving. It introduces improved mathematical reasoning capabilities through recursive proof search techniques.
What’s new in DeepSeek-Prover-V2 compared to previous versions?
The V2 version includes a more robust recursive proof search pipeline, an extended context length (up to 32K tokens in the 7B model), and stronger foundational architecture—especially in the 671B model based on DeepSeek-V3-Base.
What is recursive proof search?
Recursive proof search is a method that allows the AI to break down complex mathematical statements into smaller, more manageable sub-problems and solve them step by step, improving logical reasoning accuracy.
What is ProverBench?
ProverBench is a newly introduced benchmark dataset comprising 325 formalized math problems used to evaluate and compare the performance of neural theorem provers on a standardized set of challenges.
What kind of problems does ProverBench include?
ProverBench features 15 high-school competition-level problems from recent AIME exams and 310 problems sourced from textbooks and tutorials, covering a wide range of undergraduate-level math topics.
Why is ProverBench important?
It provides a comprehensive, diverse, and structured way to measure how well models perform across varying levels of mathematical difficulty, offering deeper insights into AI reasoning capabilities.
What model sizes are available for DeepSeek-Prover-V2?
DeepSeek-Prover-V2 comes in two variants: a 7B parameter model and a more powerful 671B parameter model for users with higher computational resources.
What is the context length supported by the 7B model?
The 7B model of DeepSeek-Prover-V2 supports up to 32,000 tokens, making it suitable for processing and understanding long, complex proofs.
What foundational models are these built on?
The 7B version builds on DeepSeek-Prover-V1.5-Base.
The 671B version is built on DeepSeek-V3-Base, offering stronger performance for large-scale tasks.
Who can benefit from using DeepSeek-Prover-V2?
Researchers in AI and mathematics, developers working on formal verification, and educators interested in AI-driven mathematical education can all benefit from using this model.
Is ProverBench publicly available?
Yes, ProverBench is part of the public release, allowing the AI research community to test and benchmark their models using the same standardized dataset.
How does this launch impact the future of AI in mathematics?
By combining powerful new models with a thoughtfully designed benchmark, DeepSeek is helping to push the boundaries of what’s possible in formal mathematical reasoning with AI—potentially transforming education, research, and automated theorem proving.
Conclusion
The launch of DeepSeek-Prover-V2 and the introduction of ProverBench represent a significant leap forward in the field of neural theorem proving. By integrating a recursive proof search pipeline and offering a diverse, structured benchmark of mathematical problems, DeepSeek AI is not only enhancing the capability of AI models to handle complex reasoning tasks but also setting a new standard for their evaluation. Whether it’s for academic research, educational use, or advancing AI in formal mathematics, this release provides the tools and framework necessary to push the boundaries of what AI can achieve in mathematical understanding and proof generation.
