今日速览
| 序号 | 标题 | 来源 | 日期 | 主题 | 推荐等级 |
|---|---|---|---|---|---|
| 1 | OpenForgeRL: Train Harness-native Agents in Any Environment | arXiv | 2026-07-23 | RAG | 高 |
| 2 | Benchmarking Agents for Proving Theorems in Quantum Algorithms and Quantum Information | arXiv | 2026-07-23 | AI-Agent | 高 |
重点论文与技术动态
1. OpenForgeRL: Train Harness-native Agents in Any Environment
- 来源：arXiv
- 日期：2026-07-23
- 作者/机构：Xiao Yu, Baolin Peng, Ruize Xu, Hao Zou, Qianhui Wu, Hao Cheng
- 主题标签：
RAG,arXiv - 推荐等级：高
- 分类：cs.AI, cs.CL
一句话结论
OpenForgeRL提供了一个开源框架，使AI代理能在任意环境中端到端训练，无需框架原生支持。
核心内容
- 解决了现有推理框架（如Codex）导致代理难以端到端训练的难题，通过解耦训练与推理过程，避免了SFT/RL堆栈无法表达有状态多进程推理的限制。
- 采用轻量级代理记录框架调用作为训练数据，并结合Kubernetes orchestrator在独立容器中运行训练回放，实现环境通用性。
- 在ClawEval、QwenClawBench和OSWorld等基准测试中，性能显著超越同类开源基线模型，如ClawEval达31.7 pass^3。
方法与数据
- 方法：轻量级代理服务框架模型调用并记录为训练数据，Kubernetes orchestrator管理容器化训练环境。
- 数据：仅需数百至数千个任务即可实现高精度训练，例如ClawEval上达31.7 pass^3。
价值判断
- 值得关注：支持研究者直接在真实部署框架（如Claude Code）和环境中训练代理，提升研究与应用的实用性。
- 可复用点：开源框架设计通用，可轻松适配任意环境和框架，无需修改底层系统。
- 局限/待核查：代理的错误恢复能力较弱，影响多步任务完成的可靠性，需进一步优化。
摘要
Modern AI agents rely on elaborate inference harnesses such as Claude Code, Codex, and OpenClaw to drive multi-turn reasoning, tool use, and access to external systems. While powerful, these complex harnesses also make agents hard to train end-to-end with open infrastructure, whose SFT/RL stacks cannot natively express stateful, multi-process harness inference. To address this, we present OpenForgeRL, an open-source framework for training harness-based agents end-to-end in diverse environments. OpenForgeRL achieves this with a lightweight proxy that serves the harness's model calls while recording them as training data for a standard RL codebase (e.g., veRL), and a Kubernetes orchestrator that runs each rollout in its own remote container, together enabling training on any harness in any environment at scale. By decoupling training and inference, OpenForgeRL allows researchers to easily train, study, and improve agents directly in the real harnesses and environments they are deployed with. We validate our framework across diverse, complex harnesses and environments, spanning tool/claw-based agents and multimodal GUI browser- and computer-use agents. Using only hundreds to a few thousand tasks, OpenForgeClaw reaches 31.7 pass^3 and 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench. OpenForgeGUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager. Both outperform open baselines of similar size on nearly all benchmarks, and in the GUI setting match or surpass models several times larger. Beyond benchmarks, we analyze how harness choice (e.g., ZeroClaw, OpenClaw, Codex) and RL shape agent behavior. We find that some harnesses are substantially harder to learn than others, and that RL improves agentic reliability, such as self-verification, tool coverage, and completing multi-step plans, though critical abilities such as error recovery remain weak.2. Benchmarking Agents for Proving Theorems in Quantum Algorithms and Quantum Information
- 来源：arXiv
- 日期：2026-07-23
- 作者/机构：Lei Zhang, Yusheng Zhao, Yimeng Cao, Ranyiliu Chen, Mingrui Jing, Jizhe Lai
- 主题标签：
AI-Agent,arXiv - 推荐等级：高
- 分类：quant-ph
一句话结论
AI代理在量子算法和信息理论定理证明中表现有限，但验证库增强可显著提升性能。
一段话。
论文构建了Lean-QuantumAlg-Bench（36任务）和Lean-QIT-Bench（40任务）两个Lean 4基准，评估了GPT-5.5等四个模型在任务基线和库增强（LAD）设置下的表现，最高难度加权分数达60.4/100和59.6/100，LAD提升性能并揭示了量子模拟等领域的系统性弱点。
核心内容
- 两个基准覆盖量子算法（36任务）和量子信息理论（40任务），任务在固定环境评估。
- LAD设置使所有模型-基准组合的分数提升最高15.9分，完成率同步提高。
- AI代理在量子模拟、量子学习、信息度量和纠缠理论中存在重复弱点。
方法与数据
- 基准任务通过确定性证明检查和目标语义审查评估，难度权重预先分配。
- 评估模型包括GPT-5.5、Kimi K3、DeepSeek V4-Pro和MiniMax M3。
价值判断
- 值得关注：基准测试为开发更可靠AI证明代理提供可重复基线。
- 可复用点：库增强推理（LAD）可有效提升领域特定证明代理性能。
- 局限/待核查：AI代理在特定量子领域表现弱；模型间成本效率差异显著。