Multi-Agent Taskforce Collaboration: Self-Correction of Compounding Errors in Long-Form Literature Review Generation

Zhi Zhang, Yan Liu, Zhejing Hu, Gong Chen, Sheng-hua Zhong, Jiannong CaoarXiv 2025 · DOI · arXiv

Introduction: The Challenge of Compounding Errors

Generating a high-quality literature review is a marathon that requires a precise sequence of tasks: searching for relevant literature, organizing a coherent outline, extracting specific evidence, and drafting thousands of words. For Large Language Model (LLM) agents, this long-horizon workflow is prone to "compounding errors." A minor mistake in the initial retrieval phase—such as selecting an irrelevant paper—cascades through the pipeline, leading to a flawed outline and a final draft plagued by hallucinations or topical drift.

Compounding errors in the literature review pipeline.

Existing self-correction methods often occur post-hoc, attempting to fix errors after the entire draft is complete. However, in long-form writing (often exceeding 4,000 words), these errors become so deeply embedded that they are nearly impossible to rectify at the final stage. Researchers identify this phenomenon as the primary barrier to autonomous scientific writing.

The SGMC Framework

To address these challenges, researchers from The Hong Kong Polytechnic University and partners have introduced Structure-Guided Memory Consolidation (SGMC). SGMC is a multi-agent framework that orchestrates a manager agent ($A_M$) and specialized worker agents (Search, Outline, Locate, Draft) to verify and consolidate information at every step of the process.

The Structure-Guided Memory Consolidation (SGMC) framework architecture.

The framework relies on three specialized memory structures to maintain grounding and logical flow:

  • Tree-Guided Memory: This module replaces flat search strategies with a hierarchical approach. It retrieves broad references, generates a sub-outline, and recursively searches for specific papers for each sub-section, ensuring the structure is rooted in actual literature.
  • Hub-Guided Memory: Serving as a central evidence "clearing house," this module uses an iterative "locate-then-draft" process. It extracts snippets, drafts sections, and re-examines literature to fill information gaps until a similarity threshold is met.
  • Self-Loop Memory: This proactive mechanism maintains a "historical experience" log. Before a task begins, the manager agent provides a checklist of past failures to avoid, mimicking human project management and preventing the repetition of known mistakes.

Evaluation and Benchmarking

The researchers evaluated SGMC using several benchmarks, including the newly introduced TopSurvey. This large-scale dataset contains 195 topics from top-tier conferences (AAAI, ACL, EMNLP) between 2023 and 2025. Because these topics are recent, they were not part of the LLMs' original training data, providing a rigorous test of the system's retrieval and synthesis capabilities.

Key Results

SGMC demonstrated state-of-the-art performance across multiple dimensions, outperforming standalone models like GPT-5.2 and Gemini-3:

  • Citation Accuracy: The system achieved a 98.17% citation recall and 89.28% precision on 8k-token benchmarks, significantly higher than previous state-of-the-art models like SurveyX.
  • Scalability: While standalone models often struggle as target lengths increase, SGMC remained stable from 8k to 64k tokens, maintaining high scores for structure and relevance.
  • Human-Level Quality: Content quality scores on the TopSurvey benchmark averaged 4.92/5.0, nearly matching human-written surveys (4.99/5.0).
  • Operational Efficiency: The framework can generate a comprehensive 8,000-token review in approximately 8.45 minutes.

Conclusion

The SGMC framework demonstrates that structure is the primary antidote to the "drift" typically seen in long-form AI generation. By implementing iterative verification and experiential checklists, SGMC moves autonomous scientific writing closer to human standards. Future work will focus on integrating multi-modal data, such as charts and tables, to further enhance the depth of generated reviews.

Summary updated

BibTeX
@misc{zhang2025multi,
  author = {Zhi Zhang and Yan Liu and Zhejing Hu and Gong Chen and Sheng-hua Zhong and Jiannong Cao},
  title = {{Multi-Agent Taskforce Collaboration: Self-Correction of Compounding Errors in Long-Form Literature Review Generation}},
  year = {2025},
  eprint = {2508.04306},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.04306},
}
Figure