Lean4AgenticRL brings formal verification to agentic RL, +20.2% on SWE-Bench
May_F1_ · x · 2026-10-25
Lean4AgenticRL: process rewards from formal verification for agentic RL
The team introduces Lean4AgenticRL, claimed as the first method applying formal verification to an open-ended domain for agentic RL:
- Lean4Agent framework: treats agent workflows as "computer programs in natural language." Its FormalAgentLib plus a precondition/postcondition predicate system verifies workflow self-consistency before execution and pinpoints which step breaks consistency for fine-grained fixes.
- Training results: using the framework's process rewards, they trained LAR-9B and LAR-27B:
- LAR-9B gains 20.2% on SWE-Bench-Verified, reaching 64.2%
- LAR-27B hits 77.3% on TerminalBench-2.1, +7.7% over the base model
- The shared benchmark table shows LAR-27B at 84.2% (SWE-Bench-Verified), 68.2% (TB-2.0), and 77.3% (TB-2.1), outperforming similar-size baselines like Qwen-3.8-27B and TMax-27B.
Paper, model weights, code, and website are all publicly available.
More from coding & agent
- Blind test of 5 local AI agents on one Qwen model: scores range from 3/5 to 5/5 — GapNew4766 · 2026-10-25
- WildAWM synthesizes open-source computer-use agent worlds; Qwen3.6 gains +10.6 across 8 benchmarks — HuaxiuYaoML · 2026-10-25
- Meme: Claude as Sisyphus endlessly pushing the 'Some Rust Port' boulder — burny_tech · 2026-10-25
- SolidCraft: Open-Source Rust Parametric CAD Nears Fusion 360 With MCP AI Agent Support — philfung · 2026-10-25
- Running Claude Code as PM over Codex workers to wire xAI Grok in as a coding agent — umeshonai · 2026-10-25
- Aaron Levie: cyber agents will soon process 1,000X more tokens as AI reshapes security — Kyrannio · 2026-10-25