Search-o1: Agentic Search-Enhanced Large Reasoning Models
Search-o1 integrates agentic retrieval and document reasoning, boosting large reasoning models' performance on complex tasks.
Key Findings
Methodology
Search-o1 employs an agentic retrieval mechanism (Agentic RAG) combined with a Reason-in-Documents module. The model autonomously generates search queries during reasoning, triggering external knowledge retrieval via vector search (e.g., FAISS). Retrieved documents undergo deep filtering through secondary generation, reducing noise and ensuring relevance. The process involves special symbols marking search requests and iterative refinement, maintaining reasoning coherence. Experiments across science, math, and coding datasets demonstrate superior performance over baseline models, with significant improvements in accuracy and robustness.
Key Results
- On GPQA science QA, Search-o1 achieved 89.4% Pass@1, outperforming the baseline (53.6%) and single retrieval models (24.1%), showing a large margin of improvement.
- In MATH500, the model reached 85% accuracy, surpassing baseline (67%), validating multi-round retrieval and filtering effectiveness.
- Across six open-domain QA benchmarks, average accuracy increased by over 12%, especially excelling in multi-hop reasoning tasks, demonstrating robustness and scalability.
Significance
This work addresses the critical bottleneck of knowledge insufficiency in large reasoning models, enabling them to dynamically access and filter external information. The approach enhances trustworthiness, interpretability, and applicability in scientific research, education, and automation. It paves the way for autonomous knowledge management in AI systems, reducing errors caused by incomplete internal knowledge and improving decision-making transparency.
Technical Contribution
The core innovation lies in integrating an agentic retrieval trigger with deep document filtering, forming a multi-round iterative reasoning framework. The method combines special token signaling, secondary generation, and vector search, enabling models to selectively incorporate external knowledge without disrupting reasoning flow. This design extends current RAG techniques, providing theoretical guarantees for relevance and coherence, and opens new avenues for scalable, explainable AI reasoning systems.
Novelty
Unlike previous static or single-shot retrieval methods, this is the first to enable models to autonomously decide when to retrieve external knowledge during reasoning. The multi-round filtering process ensures high relevance and low noise, significantly improving complex reasoning performance. This dynamic, iterative approach marks a new paradigm in knowledge-augmented AI, especially for scientific and multi-hop tasks.
Limitations
- Despite efficiency improvements, the multi-round retrieval and filtering process increases computational costs, especially with large knowledge bases.
- The deep filtering relies on pre-trained models' understanding, which may introduce biases or inaccuracies when dealing with ambiguous or incomplete documents.
- Current validation is mainly in English and Chinese datasets; cross-lingual and multi-modal reasoning capabilities need further development.
Future Work
Future directions include optimizing retrieval strategies with reinforcement learning, integrating knowledge graphs for better reasoning paths, and extending multi-modal capabilities. Improving efficiency, reducing costs, and enhancing interpretability are also priorities. Additionally, adapting the framework to real-world applications like medical diagnosis, legal analysis, and autonomous agents will be explored to realize fully autonomous, trustworthy AI systems.
AI Executive Summary
Search-o1 introduces a groundbreaking framework that combines agentic retrieval with deep document filtering to enhance large reasoning models (LRMs). Traditional models often struggle with knowledge gaps during long, multi-step reasoning, leading to errors and uncertainties. Existing retrieval methods, typically static and single-shot, cannot adequately support complex, multi-hop reasoning tasks. In response, Search-o1 employs an agentic mechanism where the model autonomously generates search queries during reasoning, triggering multi-round external knowledge retrieval via vector search tools like FAISS. Retrieved documents are then processed through a Reason-in-Documents module, which filters and condenses relevant information, ensuring high relevance and coherence. This iterative process allows the model to dynamically incorporate external knowledge without disrupting the logical flow of reasoning. Extensive experiments across scientific, mathematical, and coding datasets demonstrate that Search-o1 significantly outperforms baseline models, achieving 89.4% Pass@1 on GPQA and 85% accuracy on MATH500, among other benchmarks. The approach not only boosts accuracy but also enhances the interpretability and trustworthiness of AI reasoning systems. Its ability to adaptively access and refine external knowledge marks a major step toward autonomous, reliable AI capable of tackling real-world complex problems. Nonetheless, challenges remain in computational efficiency and multi-modal integration, which future research will address to broaden practical deployment and scalability.
Deep Dive
Abstract
Large reasoning models (LRMs) like OpenAI-o1 have demonstrated impressive long stepwise reasoning capabilities through large-scale reinforcement learning. However, their extended reasoning processes often suffer from knowledge insufficiency, leading to frequent uncertainties and potential errors. To address this limitation, we introduce \textbf{Search-o1}, a framework that enhances LRMs with an agentic retrieval-augmented generation (RAG) mechanism and a Reason-in-Documents module for refining retrieved documents. Search-o1 integrates an agentic search workflow into the reasoning process, enabling dynamic retrieval of external knowledge when LRMs encounter uncertain knowledge points. Additionally, due to the verbose nature of retrieved documents, we design a separate Reason-in-Documents module to deeply analyze the retrieved information before injecting it into the reasoning chain, minimizing noise and preserving coherent reasoning flow. Extensive experiments on complex reasoning tasks in science, mathematics, and coding, as well as six open-domain QA benchmarks, demonstrate the strong performance of Search-o1. This approach enhances the trustworthiness and applicability of LRMs in complex reasoning tasks, paving the way for more reliable and versatile intelligent systems. The code is available at \url{https://github.com/sunnynexus/Search-o1}.