Fat-Cat：面向复杂推理的文档驱动元认知多智能体系统
Fat-Cat: Document-Driven Metacognitive Multi-Agent System for Complex Reasoning
AI总结：
Fat-Cat通过文档驱动的智能体架构提升复杂推理性能，利用语义文件系统和文本策略进化模块，使Kimi-k2在HotPotQA中超越GPT-4o基线。
摘要：
LLM基于智能体的有效性往往不仅受限于模型容量本身，而是受限于运行时如何高效利用上下文信息。现有智能体框架依赖于刚性、语法重的状态表示，如嵌套JSON，这要求模型将大量注意力用于语法处理而非语义推理。本文提出Fat-Cat，一种文档驱动的智能体架构，通过改进状态管理的信噪比。通过整合三个关键组件：（1）语义文件系统，将智能体状态表示为与常见预训练语料库对齐的Markdown文档；（2）文本策略进化模块，无需参数更新即可积累任务解决知识；（3）闭环监视器，监控推理轨迹以减少幻觉。广泛的推理、检索和编码基准测试表明，Fat-Cat持续提升智能体性能。它使Kimi-k2模型在HotPotQA上超越专有GPT-4o基线。用基于文档的状态替代JSON会导致性能下降，经验验证了文档驱动状态建模相对于刚性语法的关键必要性。代码可在https://github.com/answeryt/Fat-Cat获取。
英文摘要：
The effectiveness of LLM-based agents is often limited not by model capacity alone, but by how efficiently contextual information is utilized at runtime. Existing agent frameworks rely on rigid, syntax-heavy state representations such as nested JSON, which require models to devote a substantial portion of their limited attention to syntactic processing rather than semantic reasoning. In this paper, we propose Fat-Cat, a document-driven agent architecture that improves the signal-to-noise ratio of state management. By integrating three key components: (1) a Semantic File System that represents agent state as Markdown documents aligned with common pre-training corpora, (2) a Textual Strategy Evolution module that accumulates task-solving knowledge without parameter updates, and (3) a Closed-Loop Watcher that monitors reasoning trajectories to reduce hallucinations. Extensive reasoning, retrieval, and coding benchmarks, Fat-Cat consistently improves agent performance. It enables the Kimi-k2 model to outperform the proprietary GPT-4o baseline on HotPotQA. Replacing the document-based state with JSON leads to performance drop, while empirically validating the critical necessity of document-driven state modeling over rigid syntax. The code is available at https://github.com/answeryt/Fat-Cat.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
