KPHarness: Co-Evolving Knowledge and Probing for Long-Horizon Gaming Harnesses
作者: Zhikang Chen, Yueqing Sun, Sen Cui, Ziang Ye, Yu Wang, Yan Feng2, Xierui Song, Yu Zhang, Tingting Zhu
分类: cs.AI, cs.CV
发布日期: 2026-10-20
💡 一句话要点
提出KPHarness框架以解决长时间游戏中的干预适应问题
🎯 匹配领域: 支柱九:具身大模型 (Embodied Foundation Models)
关键词: 长时间游戏 干预适应 知识探测 执行框架 动态调整 智能体训练 Minecraft任务
📋 核心要点
- 现有方法在面对失败的游戏轨迹时,缺乏有效的干预适应机制,导致无法准确调整执行框架。
- KPHarness框架通过知识探测的共同演化,结合恢复知识和探测结果,动态调整执行框架以应对失败。
- 在Minecraft任务上的实验表明,KPHarness在任务覆盖率和完成率上均优于传统的固定框架和非自适应探测方法。
📝 摘要(中文)
失败的游戏轨迹表明代理需要干预,但并不一定清楚其执行框架应如何改变。有效的框架适应需要收集诊断证据,而不仅仅是提出替代恢复规则。本文提出KPHarness框架,通过知识探测的共同演化,适应冻结的视觉-语言-动作模型的外部执行框架。恢复知识提供上下文特定的干预假设,指导主动探测;探测结果揭示这些假设何时适用,并为精炼恢复知识和后续探测选择提供证据。这一互惠过程不仅改善了应用的干预措施,也优化了调查失败的方式。
🔬 方法详解
问题定义:本文旨在解决在长时间游戏中,代理在失败轨迹下如何有效适应其执行框架的问题。现有方法往往仅依赖于固定的恢复规则,缺乏动态调整能力,导致干预效果不佳。
核心思路:KPHarness框架通过知识探测的共同演化,结合上下文特定的干预假设和探测结果,动态调整执行框架。这种设计允许框架在面对失败时,不仅能应用干预措施,还能优化调查失败的方式。
技术框架:KPHarness的整体架构包括知识探测模块、恢复知识模块和执行框架适应模块。首先,恢复知识提供干预假设,随后通过主动探测验证这些假设的有效性,最后根据探测结果调整执行框架。
关键创新:KPHarness的主要创新在于其知识探测与恢复知识的共同演化机制。这一机制使得框架能够在动态环境中自我优化,与传统方法相比,显著提高了干预的有效性和适应性。
关键设计:在实现过程中,KPHarness采用了分离的快速策略执行与缓慢的研究循环,确保了知识与探测证据的有效转化。此外,框架在每次修订时都与现有策略在匹配的沙盒条件下进行比较,确保只有在评估显示改进时才接受修订。
🖼️ 关键图片
📊 实验亮点
在Minecraft任务中的实验结果显示,KPHarness相比于传统的固定框架和非自适应探测方法,任务覆盖率显著提高,且在未见过的任务中,完成率也得到了明显改善,展示了框架的有效性和适应性。
🎯 应用场景
KPHarness框架具有广泛的应用潜力,特别是在复杂的游戏环境和机器人控制领域。通过有效的干预适应机制,该框架能够提升智能体在动态任务中的表现,未来可能在自动化游戏设计、智能机器人训练等方面发挥重要作用。
📄 摘要(原文)
A failed gaming trajectory reveals that an agent needs intervention, but not necessarily how its execution harness should change. Effective harness adaptation therefore requires gathering diagnostic evidence, rather than merely proposing alternative recovery rules. We present KPHarness, a framework that adapts the external execution harness of a frozen vision-language-action model through knowledge-probing co-evolution. Recovery knowledge supplies context-specific intervention hypotheses that guide active probing; probe outcomes reveal when these hypotheses apply and provide evidence for refining both recovery knowledge and later probe selection. This reciprocal process allows the harness to improve not only which interventions it applies, but also how it investigates failures. KPHarness separates fast policy execution from a slow research loop that translates accumulated knowledge and probing evidence into bounded harness revisions. Each revision is compared with the incumbent under matched sandbox conditions and accepted only when the paired evaluation shows improvement. Recoveries supported by further validation are retained for cross-task reuse, while immediate interaction feedback and delayed episode-success credit update action- and knowledge-probing policies during adaptation. Experiments on Minecraft tasks spanning mining, combat, and crafting show improved task coverage over bare policy execution, a manually specified fixed harness, and knowledge-only harness evolution with non-adaptive probing. Evaluation on previously unseen tasks further shows improved task completion with the evolved harness, knowledge bank, and probing policies held fixed,