Goedel-Prover-V2 argues that better training data design—not just more parameters—can move the needle in automated theorem proving. Its release spotlights two reusable patterns for applied teams: scaffolded synthetic task generation and verifier-guided self-correction.
Goedel-Prover-V2 uses scaffolded synthetic tasks and self-correction to raise theorem-proving accuracy
Goedel-Prover-V2 (released August 2025) is a series of open-source language models for automated theorem proving trained with scaffolded synthetic data—synthetic tasks generated in increasing levels of difficulty so the model learns progressively harder reasoning steps. On the MiniF2F benchmark, the 8B parameter model reports an 84.6% pass rate, outperforming the much larger 671B parameter DeepSeek-Prover-V2.
The flagship 32B model reports 88.1% pass rate in standard mode and 90.4% with verifier-guided self-correction, where the model iteratively refines outputs based on feedback. The same model is also reported to solve 86 problems on PutnamBench, substantially outpacing its competitor in the source write-up. The core claim: carefully scaffolded synthetic training tasks can close—or even invert—the usual scale advantage, producing strong results with fewer parameters.
- Cost and iteration speed: If scaffolded synthetic curricula reliably let smaller models match larger ones on narrow reasoning tasks, teams can reduce training and serving costs while iterating faster on domain-specific models.
- Data strategy becomes the lever: The approach reframes “getting more data” as “designing better synthetic tasks,” which is actionable for data leads who can control task templates, difficulty ramps, and coverage more directly than real-world data collection.
- Verifier-guided self-correction is an audit-friendly pattern: A feedback loop that forces the model to re-check and revise outputs maps well to workflows that already require validation (e.g., proofs, calculations, compliance checks), potentially improving traceability compared with single-shot generation.
- Transfer potential beyond proofs: The combination of scaffolded generation plus verification can generalize to regulated domains (legal, medical, financial) where outputs must be checked—though the verifier quality and failure modes become the critical control point.