Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fat-Cat: Document-Driven Metacognitive Multi-Agent System for Complex Reasoning

Tong Yang, Yemin Wang, Chaoning Zhang, Aming Wu

TL;DR

Fat-Cat tackles the inefficiency of runtime context in LLM-driven agents by replacing log-centric state with a Markdown-based Semantic File System, a semantic Watcher for runtime auditing, and a parameter-free Textual Strategy Evolution workflow. The methodology enforces Context Isolation to separate high-level planning from execution and uses a MemoryBridge to assemble a coherent, linear narrative, achieving $O(T)$ context growth and reducing syntactic noise. Empirical results across HotPotQA, MBPP, Bamboogle, and Med-QA show consistent gains, with up to $+14.0\%$ on Med-QA and substantial improvements over open-weight backbones that approach proprietary baselines, validating representational alignment as a key productivity lever in long-horizon reasoning. The work demonstrates that architectural choices for state representation and closed-loop control can meaningfully augment reasoning performance beyond raw parameter scaling, suggesting a robust path toward more reliable, scalable autonomous agents.

Abstract

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.

Fat-Cat: Document-Driven Metacognitive Multi-Agent System for Complex Reasoning

TL;DR

Fat-Cat tackles the inefficiency of runtime context in LLM-driven agents by replacing log-centric state with a Markdown-based Semantic File System, a semantic Watcher for runtime auditing, and a parameter-free Textual Strategy Evolution workflow. The methodology enforces Context Isolation to separate high-level planning from execution and uses a MemoryBridge to assemble a coherent, linear narrative, achieving context growth and reducing syntactic noise. Empirical results across HotPotQA, MBPP, Bamboogle, and Med-QA show consistent gains, with up to on Med-QA and substantial improvements over open-weight backbones that approach proprietary baselines, validating representational alignment as a key productivity lever in long-horizon reasoning. The work demonstrates that architectural choices for state representation and closed-loop control can meaningfully augment reasoning performance beyond raw parameter scaling, suggesting a robust path toward more reliable, scalable autonomous agents.

Abstract

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.
Paper Structure (75 sections, 15 equations, 3 figures, 6 tables, 2 algorithms)

This paper contains 75 sections, 15 equations, 3 figures, 6 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multi-Agent Frameworks & Coordination
  4. Agent Operating Systems & State Management
  5. Metacognition & Lifelong Learning
  6. Methodology
  7. Problem Formulation
  8. Stage 1: Risk-Constrained Context Initialization
  9. Stage 2: Dual-Path Strategy Evolution
  10. Stage 3: Deterministic Plan Decomposition
  11. Stage 4: Closed-Loop Control System
  12. Global Context Assembly via Document Bridge
  13. Experiments
  14. Benchmarks and Baselines
  15. Benchmarks.
  16. ...and 60 more sections

Figures (3)

  • Figure 1: Overview of the Fat-Cat Architecture. The system acts as a bicameral operating system: the Cognitive Wing (left) handles high-level strategy retrieval and risk assessment, while the Executive Wing (right) manages deterministic planning and closed-loop monitoring. The Watcher acts as a semantic firewall, auditing tool outputs against the step.md plan before committing to the global Markdown state.
  • Figure 2: Complexity-Performance Correlation. Regression analysis reveals a strong positive correlation ($R^2 = 0.98$) between task complexity (CCI) and performance gain. This confirms that the document-driven architecture provides benefits as cognitive load increases, effectively amortizing the syntactic tax.
  • Figure 3: Qualitative analysis of Fat-Cat's control mechanisms on a GAIA case study. (1) Strategy Evolution: The agent combines previously separate reasoning patterns into a new heuristic ($S_{15}$), which guides planning before execution. (2) Representational Alignment: State is maintained in a Markdown-based semantic file system, reducing syntactic overhead and preserving salient information in the context. (3) Semantic Firewall: An independent Watcher monitors intermediate actions and blocks invalid operations, preventing error propagation during execution.