DeepSeek-AI Open-Sources DeepSeek-Prover-V1.5: A Language Model with 7 Billion Parameters that Outperforms all Open-Source Models in Formal Theorem Proving in Lean 4

DeepSeek-Prover-V1.5: Advancing Formal Theorem Proving

Practical Solutions and Value

DeepSeek-Prover-V1.5 introduces a unified approach for formal theorem proving, addressing challenges faced by large language models (LLMs) in mathematical reasoning and theorem proving using systems like Lean and Isabelle.

Key Highlights:

Enhanced base model with further training on mathematics and code data, focusing on formal languages like Lean, Isabelle, and Metamath.
Improved Lean 4 code completion dataset through data augmentation techniques.
Utilized reinforcement learning from proof assistant feedback and advanced tree search methods.

Significant Advancements:

DeepSeek-Prover-V1.5-RL achieved a 60.2% pass rate in whole-proof generation, marking a 10.2 percentage point improvement over its predecessor.
On the miniF2F-test dataset, it proved 51.6% of problems with a limited sampling budget of 128 attempts, outperforming other methods.
DeepSeek-Prover-V1.5-RL achieved a state-of-the-art 62.7% pass rate with RMaxTS tree search.
Outperformed existing methods on the ProofNet dataset, demonstrating superior performance across different theorem-proving tasks and methodologies.

Key Features:

7 billion parameter language model
Specialized pre-training, supervised fine-tuning, and reinforcement learning via GRPO
Incorporates RMaxTS, an innovative Monte-Carlo tree search variant

Future Developments:

While the current focus is on exploration, future developments may include a critic model for assessing incomplete proofs, addressing the exploitation aspect of reinforcement learning in theorem proving.

Stay Connected:

Check out the Paper and GitHub
Follow on Twitter, join the Telegram Channel, and connect on LinkedIn

Evolve Your Company with AI

Discover how AI can redefine your way of work. Identify Automation Opportunities, Define KPIs, Select an AI Solution, and Implement Gradually. For AI KPI management advice, connect with us at hello@itinai.com. Stay tuned on Telegram or Twitter for continuous insights into leveraging AI.