Search-o1: Agentic Search-Enhanced Large Reasoning Models

> Source: arXiv:2501.05366 | Code: github.com/sunnynexus/Search-o1 | Posted 2025-01-09 | Authors: Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, et al. (8 authors)

Overview

Large reasoning models (LRMs) like OpenAI-o1 demonstrate impressive long stepwise reasoning via large-scale reinforcement learning, but their extended reasoning often suffers from knowledge insufficiency, leading to uncertainty and errors. Search-o1 addresses this with two components:

1. Agentic RAG mechanism — an agentic search workflow is integrated into the reasoning process, triggering dynamic retrieval of external knowledge whenever the LRM hits uncertain knowledge points. 2. Reason-in-Documents module — since retrieved documents are verbose, a separate module deeply analyzes the retrieved information before injecting it into the reasoning chain, minimizing noise and preserving coherent reasoning flow.

Original Abstract (English)

> 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. 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.

Key Points

• Positioning: sits at the intersection of agentic search and large-scale search/recommendation systems — rethinking the division of labor between retrieval, ranking, generation, and tool calling in the LLM era. The new variables are reasoning budget and action space (whether to retrieve, how many times, which tools to call).
• Evaluation: tested on complex reasoning in science, math, and coding plus six open-domain QA benchmarks; consult the PDF tables for exact numbers.
• Related work trail: the forum post cross-references RAG surveys and agentic-IR literature, e.g. Agentic Information Retrieval (arXiv:2410.09713) and a systematic review of synergizing RAG and reasoning (arXiv:2504.15909).

Engineering Takeaways

• Architecture: cascade retrieval + rerank + generation remains mainstream, but the agentic paradigm makes retrieval count and strategy themselves learnable/decision points.
• Latency & cost: budget p99 latency and per-query token spend; use early stopping, caching, and iterative-retrieval caps.
• Safety: open-web retrieval can introduce poisoning/bias; use source whitelists and output filtering.
• Evaluation gap: offline nDCG-style metrics diverge from online satisfaction; validate LLM-as-judge against human audits and measure citation accuracy and multi-hop chain completeness.

Glossary

| Term | Meaning | |------|---------| | RAG | Retrieval-Augmented Generation | | LRM | Large Reasoning Model | | Agentic Search | Search modeled as sequential decision-making and tool calling | | nDCG | Normalized Discounted Cumulative Gain, a ranking quality metric |

Notes

Specific numeric results should be verified against the original paper's tables; this post summarizes based on the abstract and public metadata. Reference readers to the paper and official code repository for reproduction.