Search New Paradigm! AI Search Paradigm Redefines Intelligent Search for Complex Information Needs}

Recently, Baidu’s search team published a paper titled "Towards AI Search Paradigm", which is the first comprehensive technical report proposing the full concept of AI search, detailing the algorithms, and demonstrating its deployment in a search scenario with hundreds of millions of daily active users.

This AI search paradigm aims to simulate human information processing and decision-making, using a modular system driven by multiple large language model (LLM) agents. It can adapt dynamically from simple fact queries to complex multi-stage reasoning tasks. These agents collaborate through dynamic workflows, evaluate query complexity, and under the guidance of a Master agent, assemble different specialized agents. They decompose complex problems into executable tasks and coordinate tool usage for task execution and content synthesis. The technical details include multi-agent collaboration, task planning, model context protocol (MCP) for tool selection, DeepResearch, reinforcement learning-based optimization, retrieval and ranking strategies, retrieval-augmented generation (RAG) alignment, and lightweight inference acceleration algorithms.

Paper link: https://arxiv.org/pdf/2506.17188

1. Research Problem: Limitations of Traditional Search in Complex Reasoning

From lexical search to machine learning-based search, and now RAG, search technology has continuously advanced. Despite RAG systems’ ability to provide answers in natural language, they struggle with multi-stage reasoning, complex task decomposition, or cross-tool collaboration. The core issue is that traditional RAG relies on a linear "retrieve-then-generate" process, lacking mechanisms for cognitive planning, dynamic tool invocation, reflection, and adjustment, making it difficult to handle complex, diverse information needs. For example, RAG can answer "What is Emperor Wu of Han’s age?" but cannot easily answer "Who is older, Emperor Wu or Caesar, and by how many years?" which requires complex reasoning.

The fundamental reason is that most RAG systems are still linear, single-round retrieval and generation, lacking task decomposition, tool scheduling, information validation, error reflection, and answer integration. To address this, Baidu’s team proposes a new AI search paradigm that mimics human information acquisition, like assembling a team of experts within the search engine: it can automatically decompose tasks, invoke multiple tools, and use multi-agent collaboration to solve complex problems.

2. Method Introduction: A Modular, Multi-Agent Collaborative Search Paradigm

The paper introduces the AI Search Paradigm, a new modular, multi-agent collaboration-based search framework. Its core idea is to enable multiple dedicated agents to dynamically cooperate, achieving efficient planning, reliable execution, and result integration for complex queries. It includes four agent roles:

• Master (Control Agent): Analyzes query complexity, dynamically assembles the optimal agent team, monitors progress, and adjusts strategies based on feedback.
• Planner (Task Planning Agent): Adjusts LLM capabilities based on query complexity, constructs directed acyclic graphs (DAG), and dynamically calls relevant tools.
• Executor (Task Execution Agent): Executes sub-tasks, invoking external tools like web search, calculators, weather APIs, code interpreters, and evaluates execution results in real-time.
• Writer (Answer Generation Agent): Integrates sub-task results to generate clear, coherent natural language answers.

This paradigm breaks the traditional linear, static search process by introducing explicit task planning, dynamic tool invocation, and real-time reflection, significantly enhancing the system’s ability to handle complex tasks.

^{Figure 1. Architecture diagram of the AI Search Paradigm system.}

3. Method Design: From Static Retrieval to Dynamic Collaboration, Building a More "Human-like" Search Workflow

The AI Search Paradigm is not just about adding modules but involves a complete architectural overhaul:

For example, a complex query like:

"Who is older, Emperor Wu of Han or Caesar? By how many years?"

The AI Search system plans the following task graph:

Sub-task 1: Retrieve Emperor Wu’s birth date

Sub-task 2: Retrieve Caesar’s birth date

Sub-task 3: Calculate age difference

Dependencies exist between sub-tasks (e.g., calculating age requires obtaining birth dates first). The system constructs a DAG and executes tasks in layers and parallel. Compared to the linear "think-action" RAG, AI Search is more stable, scalable, and traceable.

^{Figure 2. Comparison between AI Search Paradigm and mainstream RAG methods.}

4. Key Technical Highlights

4.1 Master

• Adjusts the team dynamically based on query complexity
• Monitors and evaluates task results in real-time, initiating reflection (Reflect), re-planning (Replan), and re-execution (Re-Act) as needed

4.2 Planner

• Dynamic Capability Boundary: Unlike static tool lists, it dynamically selects appropriate tools based on input query, combining LLM capabilities with tool APIs to form effective capability boundaries.

^{Figure 3. Illustration of dynamic capability boundary.}

• Constructing Accurate and Complete Tool Description Documents (DRAFT): Uses interaction feedback between LLM and external tools to iteratively optimize tool descriptions.

^{Figure 4. Architecture of the DRAFT tool description construction algorithm.}

• Tool Clustering in MCP: Clusters tools based on API similarity to enhance system robustness through redundancy.
• Query-based Tool Retrieval (COLT): Supports semantic-based tool selection, relationship modeling, and multi-label ranking to ensure complete and rational tool combinations.
• Task Planning Based on DAG: Decomposes complex problems into sub-tasks, constructs JSON-formatted DAGs, and performs internal reasoning before generating structured DAGs.
• DeepResearch under Master Guidance: Executes each step with validation; if failure occurs, triggers reflection, re-planning, and local DAG rollback to avoid full re-computation.
• RL-based Optimization of Planner: Uses reinforcement learning with explicit reward mechanisms (accuracy, user feedback, format, intermediate task quality) to improve planning performance.

4.3 Executor

• Aligns with LLM Preferences: Focuses on aligning with large model preferences rather than just user preferences, to generate higher-quality answers.

^{Figure 5. Architecture transition of the task executor.}

• LLM Annotation (RankGPT and TourRank): Uses sliding window comparisons and tournament-style multi-round scoring to generate more accurate document rankings.
• Reference Selection and Generation Rewards: Upgrades from single-dimension ranking to multi-dimensional strategies, with the system generating answers based on different strategies and using answer quality to drive optimization.

^{Figure 6. Illustration of answer generation reward algorithms.}

• LLM Distillation: Distills deep ranking ability from teacher LLM into a lightweight student model.
• Lightweight Retrieval and Ranking System: Uses Tiny-LLM for query and document embedding generation, replacing traditional BERT models, and directly scores documents with LLM inference to reduce feature engineering.

4.4 Writer

• Based on the 3H Standard: Helpfulness, Harmlessness, and Honesty, with robust RAG systems (ATM), multi-agent iterative training, and explicit user feedback for alignment and robustness.

^{Figure 7. Illustration of the Robust RAG algorithm (ATM).}

^{Figure 8. Illustration of the RAG task alignment algorithm (PA-RAG).}

• Multi-Agent Joint Optimization (MMOA-RAG): Treats Planner, Executor, and Writer as a collaborative system, optimized via multi-agent PPO (MAPPO) with penalties to prevent inefficiency.

^{Figure 9. Multi-agent joint optimization algorithm (MMOA-RAG).}

4.5 Lightweight LLM

• Algorithmic Optimization: Proposes Local Attention, replacing global attention with local or low-rank mechanisms, reducing complexity to near-linear with minimal performance loss; uses pruning techniques like Layer Collapse and SlimGPT to compress models without retraining.
• Architectural Optimization: Implements output length reduction, semantic caching, quantization, prefill-decoding separation, and speculative decoding for lightweight deployment.

^{Figure 10. Flowchart of lightweight LLM technology.}

5. Experimental Results: Leading Performance of the AI Search System

The rigorous evaluation confirms the system’s powerful capabilities.

5.1 Human Evaluation

Blind tests by professional annotators show that for simple queries, AI search and traditional search perform similarly; for moderate queries, AI search outperforms by 5%; and for complex queries, the advantage increases to 13%.

Table 1. Comparison of AI search and traditional search in human evaluation experiments.

5.2 A/B Testing

Using real traffic data from Baidu Search, results show AI search significantly improves key user metrics: query change rate (CQR) decreased by 1.45%, page views (PV) increased by 1.04%, daily active users (DAU) increased by 1.85%, and dwell time increased by 0.52%.

Table 2. Online A/B test results.

5.3 Typical Case Analysis

Two cases are highlighted:

• Simple Query: "How tall is Mount Tai?" — both AI and traditional web search perform similarly.
• Complex Query: "Who is older, Emperor Wu of Han or Caesar? And by how many years?" — AI search provides accurate answers, while traditional systems fail.

Figure 11. Case analysis.

Future Outlook of AI Search

The AI Search Paradigm offers a new framework that transforms traditional retrieval-generation mechanisms into a cognitive intelligent search system centered on multi-agent collaboration, task planning, and dynamic tool invocation. It is expected to bring profound technological advances to search engines, intelligent Q&A, knowledge base interactions, and decision support systems, truly entering a new era of intelligent search.

The paper is publicly available on Arxiv: https://arxiv.org/pdf/2506.17188. More AI search experiences can be explored at: https://www.baidu.com/.