Fat-Cat: Document-Driven Metacognitive Multi-Agent System for Complex Reasoning
This addresses a bottleneck in multi-agent systems for complex reasoning, offering a novel approach to reduce syntactic overhead and enhance semantic processing.
The paper tackles the problem of inefficient contextual information utilization in LLM-based agents by proposing Fat-Cat, a document-driven architecture that improves state management, resulting in performance gains such as enabling the Kimi-k2 model to outperform GPT-4o on HotPotQA.
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.