Recently, the Seed team at ByteDance launched its latest formal mathematical reasoning model, Seed Prover 1.5, marking a significant breakthrough in the field of mathematical reasoning. This release was made possible by in-depth exploration of large-scale Agentic reinforcement learning, significantly improving its reasoning capabilities and efficiency.
During the 2025 International Mathematical Olympiad (IMO), Seed Prover demonstrated its powerful performance. Within just three days, the model successfully solved four out of six problems and provided partial proofs for one, ultimately earning an official silver medal. In comparison, Seed Prover 1.5 generated complete, compilable, and verifiable Lean proof code for the first five problems of IMO2025 within 16.5 hours, meeting the standard for a gold medal.

More notably, Seed Prover 1.5 also performed outstandingly in the 2025 Putnam Mathematical Competition, completing 11 out of 12 problems by generating compilable and verifiable Lean code within 9 hours. This achievement set new records for formal mathematical reasoning models across multiple evaluation sets, especially in assessments with math difficulty levels comparable to those of master's and doctoral students, where it solved 80% and 33% of the problems respectively.
The innovation of Seed Prover 1.5 lies in its new Agentic Prover architecture, which combines the advantages of natural language reasoning and formal proofs. Unlike previous formal provers, Seed Prover 1.5 can flexibly call various tools during the proof process, such as actively retrieving the extensive mathematics library Mathlib and executing Python code to assist verification. Through incremental lemma verification, the model breaks down complex problems into several lemmas and gradually builds formal proofs.
Additionally, Seed Prover 1.5 introduced the Sketch Model, which simulates the problem-solving approach of human mathematicians, transforming natural language proofs into high-level proof frameworks, thus significantly reducing the difficulty of proving complex theorems. By adopting this "divide and conquer" strategy, Seed Prover 1.5 can effectively avoid error accumulation when generating long texts.
Technical Report:
https://arxiv.org/abs/2512.17260
Lean Proof Code:
https://github.com/ByteDance-Seed/Seed-Prover/blob/main/SeedProver-1.5/Putnam2025.zip
