---
id: 20261020-T0-01
title: "Lean4AgenticRL：用可验证的多级过程奖励提升LLM智能体能力"
title_en: "Lean4AgenticRL: Verifiable Multi-level Process Rewards Boost LLM Agents"
url: https://ai.daily.yangsir.net/daily/20261020-T0-01
issue_date: 2026-10-20
publish_date: 2026-10-19T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.22448
---

# Lean4AgenticRL：用可验证的多级过程奖励提升LLM智能体能力

arXiv新论文Lean4AgenticRL提出用可验证的多级过程奖励推进智能体强化学习，目标场景是复杂软件工程和通用终端操作。当前顶级LLM在这类任务上依赖大规模智能体强化学习，但过程奖励难以验证。该工作把奖励拆成多级并保证可验证性，若效果成立，可减少对结果奖励的依赖，让训练信号更密集，对SWE类任务的智能体训练有直接参考价值。

## English Version

**Lean4AgenticRL: Verifiable Multi-level Process Rewards Boost LLM Agents**

A new arXiv paper, Lean4AgenticRL, proposes verifiable multi-level process rewards for agentic reinforcement learning, targeting complex software engineering and general terminal-use tasks. State-of-the-art LLMs rely on large-scale agentic RL for these tasks, but process rewards are hard to verify. This work splits rewards into multiple levels while keeping them verifiable; if the results hold, it reduces reliance on outcome rewards and densifies the training signal, offering a direct reference for agent training on SWE-style tasks.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.22448)

**详情页**：https://ai.daily.yangsir.net/daily/20261020-T0-01

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*