OmniThink: Expanding Knowledge Boundaries in Machine Writing through Thinking

Xi Zhang, Wenbiao Yin, J. Fang, Jing Wu, Runnan Fang, Yong Jiang, Pengjun Xie, Fei Huang, Huajun Chen, Ningyu Zhang2025 · DOI

In the current landscape of Large Language Models (LLMs), generation is often characterized by "fast thinking"—the immediate production of a response upon receiving a prompt. While Retrieval-Augmented Generation (RAG) has improved factual grounding, long-form outputs frequently remain shallow, repetitive, or constrained by the model's initial training distribution.

A recent study by researchers from Zhejiang University and Alibaba Group, titled "OmniThink: Expanding Knowledge Boundaries in Machine Writing through Thinking," introduces a paradigm shift toward "slow thinking." By simulating human-like iterative research and reflection, OmniThink aims to transcend the "knowledge boundaries" that typically limit machine-generated content.

The Problem: Knowledge Boundaries in Machine Writing

The authors identify two primary bottlenecks in existing machine writing systems:

  1. The Knowledge Information Boundary: Models often struggle to access diverse or deep information, frequently relying on a static, narrow set of search results.
  2. The Knowledge Cognition Boundary: Even when high-quality information is retrieved, models may fail to organize it effectively. This leads to high redundancy—often referred to as "fluff"—where the same core facts are repeated throughout an article without adding new value.

While role-playing frameworks (like STORM) attempt to broaden perspectives, they remain tethered to predefined "persona" scopes. OmniThink moves beyond these constraints by treating writing as a continuous, self-directed cycle of learning and sense-making.

Methods: Simulating Human Cognitive Behavior

OmniThink utilizes two core components to emulate human research processes:

  • The Information Tree ($\mathcal{T}$): A hierarchical structure where the root represents the main topic and branches denote increasingly granular subtopics. Rather than a single search pass, the model analyzes its current tree to identify "knowledge gaps" and generates targeted queries to expand into new domains.
  • The Conceptual Pool ($\mathcal{P}$): This serves as the model's evolving internal representation. As new data is retrieved, the model reflects on it—filtering, distilling, and synthesizing insights. This pool then informs subsequent expansion phases, ensuring the model "understands" the context before seeking more data.
The OmniThink workflow: Iterative expansion and reflection to build an information tree and conceptual pool.

The workflow consists of four stages:

  1. Initialization: Establishing a root node and an initial conceptual pool.
  2. Iterative Expansion & Reflection: Growing the Information Tree while concurrently updating the Conceptual Pool.
  3. Concept-guided Outline Structuring: Using distilled insights from the pool to refine a logical, comprehensive outline.
  4. Article Composition: Generating sections in parallel using documents associated with specific branches of the Information Tree.

Evaluation and Key Results

The framework was evaluated using the WildSeek dataset, comprising 100 complex topics across 24 domains. To quantify quality, the researchers introduced Knowledge Density (KD), a metric measuring the ratio of unique "atomic knowledge units" to total text length, effectively penalizing verbosity and rewarding information richness.

OmniThink was benchmarked against RAG, oRAG, and STORM using various LLM backbones (GPT-4o, Qwen-Plus, O1-preview, and DeepSeek-R1).

Performance comparison: OmniThink significantly outperforms baselines across relevance, breadth, depth, and novelty.

Key findings include:

  • Superior Novelty and Depth: OmniThink consistently outperformed STORM and Co-STORM, particularly in "Novelty" (e.g., scoring 4.31 vs. STORM's 3.80 on GPT-4o).
  • Higher Knowledge Density: The system achieved the highest KD scores, providing more substantive information per paragraph.
  • Information Diversity: OmniThink retrieved an average of 120.63 unique URLs per topic, a massive increase over STORM's 16.56.
  • Human Preference: Human evaluators reported an 11% improvement in breadth over the strongest baselines.

Conclusion and Future Directions

The authors conclude that expansion drives depth and breadth, while reflection is essential for novelty and reducing redundancy. By allowing a model the "time" to iteratively refine its internal representation, OmniThink achieves a qualitative leap from simple search summarization to expert-level report synthesis.

Future work will focus on expanding the framework to include multimodal data (images, video) and exploring ways to adapt the system's academic tone to more diverse or creative writing styles.

Citation: Zekun Xi, Wenbiao Yin, Jizhan Fang, Jialong Wu, Runnan Fang, Yong Jiang, Pengjun Xie, Fei Huang, Huajun Chen, Ningyu Zhang. "OmniThink: Expanding Knowledge Boundaries in Machine Writing through Thinking." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 2025.

Summary updated

BibTeX
@misc{zhang2025omnithink,
  author = {Xi Zhang and Wenbiao Yin and J. Fang and Jing Wu and Runnan Fang and Yong Jiang and Pengjun Xie and Fei Huang and Huajun Chen and Ningyu Zhang},
  title = {{OmniThink: Expanding Knowledge Boundaries in Machine Writing through Thinking}},
  year = {2025},
  doi = {10.18653/v1/2025.emnlp-main.50},
}
Figure