Lean4AgenticRL: Advancing Agentic RL with Verifiable Multi-level Process Rewards

Benchmark Model Rank Results
code-generation-on-terminal-bench-2-0-terminus-2LAR-27B (Terminus-2)#1Accuracy: 68.2
code-generation-on-swe-bench-verifiedLAR-27B (Claude-code)#4Accuracy: 84.2
code-generation-on-terminal-bench-2-1LAR-27B (Terminus-2)#20Accuracy: 77.3