[论文] AgentSpec: Understanding Embodied Agent Scaffolds Through Controlled C...

论文概要

研究领域: CL
作者: Jixuan Chen, Jianzhi Shen, Haoqiang Kang
发布时间: 2026-06-12
arXiv: 2606.14674

中文摘要

LLM智能体越来越多地不是作为单一模型调用构建，而是作为结合推理、记忆、反思、动作执行和学习的脚手架系统构建。虽然这些脚手架通常提高性能，但它们通常嵌入紧密耦合的管道中，使得隔离组件贡献、比较替代设计或理解模块交互如何塑造智能体行为变得困难。我们引入了AgentSpec，一个模块化规范框架，将具身智能体表示为可重用策略组件的类型化组合，具有标准化接口。AgentSpec标准化了感知、记忆、推理、反思、动作和可选学习之间的接口，使组件能够在受控条件下被交换和重新组合。我们在DeliveryBench、ALFRED、MiniGrid和RoboTHOR上实例化该框架，并分析跨模型主干的推理、记忆、反思和强化学习模块。我们的结果表明，智能体性能由脚手架兼容性和交互效应而非孤立的模块强度决定。特别是，结构化多粒度记忆改善长程状态跟踪，推理和记忆在不同环境中非均匀交互，反思在纠正和成本之间权衡，而RL训练的策略在与部署时脚手架结构优化时组合最佳。AgentSpec为研究、比较和设计可组合LLM智能体提供了受控基础。我们的代码、基线和交互式游乐场在 https://agentspec-embodied.github.io 公开可用。

原文摘要

LLM agents are increasingly built not as single model calls, but as scaffolded systems that combine reasoning, memory, reflection, action execution, and learning. While such scaffolds often improve performance, they are often embedded in tightly coupled pipelines, making it difficult to isolate component contributions, compare alternative designs, or understand how module interactions shape agent behavior. We introduce AgentSpec, a modular specification framework that represents embodied agents as typed compositions of reusable policy components with standardized interfaces. AgentSpec standardizes the interfaces among perception, memory, reasoning, reflection, action, and optional learning, enabling components to be swapped and recombined under controlled conditions. We instantiate this fr...

自动采集于 2026-06-16

#论文 #arXiv #CL #小凯