Lean Theorem Proving on PROOFNET (186 problems)
24.73Pass@8Post-hoc Agentic SFT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Post-hoc Agentic SFTBase checkpoint=GOEDEL-PROVER-V2-32B-SFT, Lean agentic SFT data size=18K2026.04 | 24.73 | 27.96 | 88.5 | |
| Post-hoc Agentic SFTBase checkpoint=GOEDEL-PROVER-V2-32B-RL, Lean agentic SFT data size=1002026.04 | 24.73 | 26.88 | 90 | |
| Post-hoc Agentic SFTBase checkpoint=GOEDEL-PROVER-V2-32B-SFT, Lean agentic SFT data size=1K2026.04 | 23.66 | 27.96 | 92.3 | |
| Post-hoc Agentic SFTBase checkpoint=GOEDEL-PROVER-V2-32B-RL, Lean agentic SFT data size=1K2026.04 | 23.12 | 26.88 | 94 | |
| Post-hoc Agentic SFTBase checkpoint=GOEDEL-PROVER-V2-32B-RL, Lean agentic SFT data size=18K2026.04 | 21.51 | 26.88 | 90 | |
| Post-hoc Agentic SFTBase checkpoint=GOEDEL-PROVER-V2-32B-SFT, Lean agentic SFT data size=1002026.04 | 20.97 | 25.81 | 93.8 | |
| GOEDEL-PROVER-V2-32B-SFTCheckpoint Type=SFT checkpoint2026.04 | 17.74 | 21.51 | 0 | |
| GOEDEL-PROVER-V2-32B-RLCheckpoint Type=Post-RL checkpoint2026.04 | 16.67 | 22.58 | 0 | |
| QWEN3-32B2026.04 | 8.6 | 13.44 | 96 |