Search-o1: Agentic Search-Enhanced Large Reasoning Models
Overview
- Paper: Search-o1: Agentic Search-Enhanced Large Reasoning Models
- Authors: Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, et al. (8 authors total)
- Date: 2025-01-09
- Code: https://github.com/sunnynexus/Search-o1
- Category: Agentic Search
- Tasks: Complex reasoning in science, mathematics, and coding, plus six open-domain QA benchmarks.
- Baselines: Non-retrieval LRMs, standard RAG pipelines, BM25/dense retrieval variants.
- Findings: Search-o1 shows strong performance across benchmarks, improving trustworthiness and applicability of LRMs in knowledge-intensive reasoning. Exact figures should be verified against the original PDF tables.
- Synergizing RAG and Reasoning: A Systematic Review (arXiv:2504.15909)
- Agentic Information Retrieval (arXiv:2410.09713)
- AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforcement Learning
Motivation
Large reasoning models (LRMs) like OpenAI-o1 demonstrate impressive long stepwise reasoning through large-scale reinforcement learning. However, their extended reasoning processes often suffer from knowledge insufficiency, leading to frequent uncertainties and potential errors. Traditional pipelines also separate retrieval, ranking, and generation, making them poorly suited to the natural-language, multi-hop, and real-time knowledge demands of the LLM era. Search-o1 addresses this gap by tightly integrating agentic retrieval into the reasoning process itself.
Method
Search-o1 combines two key components:
1. Agentic RAG mechanism: An agentic search workflow is embedded into the reasoning process. When the LRM encounters an uncertain knowledge point, it dynamically triggers retrieval of external knowledge instead of relying on parametric memory alone.
2. Reason-in-Documents module: Because retrieved documents are verbose and noisy, a separate module deeply analyzes the retrieved information before injecting it into the reasoning chain — minimizing noise and preserving a coherent reasoning flow.
The broader design follows the typical agentic-search pattern: encode queries and documents into representations, compose core modules (retriever, reranker, planner, generator, tool interfaces), then apply iterative retrieval with early stopping and budget control during inference.
Evaluation
Experiments cover:
Key Takeaways
1. Architecture: Agentic paradigms make retrieval frequency and strategy themselves learnable/adaptive decisions rather than fixed pipeline stages. 2. Reasoning quality: Deep document analysis (Reason-in-Documents) before injection is critical to avoid context pollution in long reasoning chains. 3. Evaluation: The field is shifting from static nDCG toward task success rate, citation accuracy, and multi-hop reasoning chain completeness. 4. Deployment: Latency, cost, and safety (e.g., retrieval poisoning, hallucination) remain hard constraints for industrial adoption.
Limitations and Future Work
Likely limitations include experiment scale constrained by GPU budget, benchmark-vs-real-user distribution mismatch, English-centric data limiting cross-language generalization, and safety risks of open-web agentic retrieval. Future directions include more efficient test-time compute allocation, deeper integration with knowledge graphs, and causal/fairness constraints for recommendation systems.
Original Abstract (verbatim)
> 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 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. [...] The code is available at https://github.com/sunnynexus/Search-o1.