Lean4AgenticRL trains LAR-9B and LAR-27B with process rewards
Original title: Lean4AgenticRL: Advancing Agentic RL with Verifiable Multi-level Process Rewards
Authors report LAR-27B at 77.3% on TerminalBench-2.1 and a 9.89% average gain over the base model.
The paper proposes Lean4AgenticRL, which grounds multi-level process rewards in a Lean4 schema for open-ended agentic reinforcement learning on software engineering and terminal tasks. It builds problem-specific workflows, verifies tool-call and workflow-stage behavior, and trains with workflow-stage-normalized GRPO. The authors train LAR-9B and LAR-27B and say the models beat similar-size leaders on their tests.
Unconfirmed: Benchmark gains are author-reported; independent verification is not stated.
- Rewards cover tool calls, workflow stages, and outcomes.
- Training uses workflow-stage-normalized GRPO.
- Authors claim a 5.07% edge over the second-best comparable model.
- The trained family is LAR-9B and LAR-27B.
From the original article
1 min read at huggingface.co
Enhancing the agentic capabilities of Large Language Models (LLMs) to tackle complex software engineering (SWE) and general terminal-usage problems is a pressing challenge. Current state-of-the-art LLMs often go through large-scale agentic Reinforcement Learning (RL).
By Ruida Wang, Gerald Friedland, Tong Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.