Source-linked AI summary
Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization
Qianben Chen, Tianrui Qin, King Zhu, Qiexiang Wang, Chengjun Yu, Shu Xu, Jiaqi Wu, Jiayu Zhang, Xinpeng Liu, Xin Gui, Jingyi Cao, Piaohong Wang, Dingfeng Shi, He Zhu, Tiannan Wang, Yuqing Wang, Maojia Song, Tianyu Zheng, Ge Zhang, Jian Yang, Jiaheng Liu, Minghao Liu, Yuchen Eleanor Jiang, Wangchunshu Zhou
TL;DR
Long-horizon agentic search faces high latency from deeper sequential reasoning and weak generalization across heterogeneous task types. SMTL uses parallel evidence acquisition, structured context management, and unified multi-type data synthesis, achieving strong benchmark performance while reducing reasoning effort.
Problem
Longer reasoning trajectories increase inference cost and latency, while agentic search generalization across deterministic and open-ended research settings remains challenging.
Method
SMTL replaces sequential reasoning with parallel evidence acquisition and structured context management, and synthesizes tasks across deterministic question answering and open-ended research.
Results
SMTL achieves state-of-the-art or competitive performance across deep search and research benchmarks while substantially reducing reasoning steps and inference latency.
Takeaways & Limitations
Efficient, search-centric scaling provides a practical and generalizable foundation for future deep research agents.
Abstract
from arXiv · showhide
Recent deep research agents primarily improve performance by scaling reasoning depth, but this leads to high inference cost and latency in search-intensive scenarios. Moreover, generalization across heterogeneous research settings remains challenging. In this work, we propose \emph{Search More, Think Less} (SMTL), a framework for long-horizon agentic search that targets both efficiency and generalization. SMTL replaces sequential reasoning with parallel evidence acquisition, enabling efficient context management under constrained context budgets. To support generalization across task types, we further introduce a unified data synthesis pipeline that constructs search tasks spanning both deterministic question answering and open-ended research scenarios with task appropriate evaluation metrics. We train an end-to-end agent using supervised fine-tuning and reinforcement learning, achieving strong and often state of the art performance across benchmarks including BrowseComp (48.6\%), GAIA (75.7\%), Xbench (82.0\%), and DeepResearch Bench (45.9\%). Compared to Mirothinker-v1.0, SMTL with maximum 100 interaction steps reduces the average number of reasoning steps on BrowseComp by 70.7\%, while improving accuracy.
1 Introduction
SMTL targets the efficiency and generalization limits of long-horizon agentic search by replacing sequential reasoning with parallel evidence acquisition and broadening task construction across deterministic and open-ended settings.
- Longer reasoning trajectories improve search performance but increase inference latency, leaving efficiency unresolved.
- Generalization remains challenging across deterministic question-answering tasks and open-ended research tasks with different evaluation criteria.
- SMTL replaces sequential reasoning with parallel evidence acquisition and structured context management under constrained context budgets.
- Its automated data pipeline constructs representative multi-type search tasks spanning deterministic and open-ended settings.
- SMTL achieves state-of-the-art or competitive performance while substantially reducing reasoning steps and inference latency.
2 Related Work
Existing agentic workflows commonly scale search performance through deeper sequential reasoning, while synthetic data pipelines face costly collection and limited task coverage.
- Agentic workflows augment language models with planning, multi-step tool use, and iterative environment interaction for search-intensive tasks.
- Most existing workflows deepen sequential reasoning and expand interaction horizons, often limiting information efficiency.
- Collecting multi-step, tool-interactive trajectories at scale remains costly for search-agent training.
3 Parallel Agentic Workflow
The parallel agentic workflow decomposes composite search tasks into concurrent subtasks, aggregates their observations, and periodically refines the plan during execution.
- The agent initializes a task plan, executes multiple subtasks in parallel, and synchronizes results through periodic plan refinement.
- Initial planning decomposes a composite problem into interrelated, partially independent subtasks with parallelizable execution paths.
- Pending subtasks are processed concurrently, and their actions and observations are aggregated into a unified reasoning state.
- Parallel execution uses reusable external tools, including web search and page crawling, for concurrent information acquisition and verification.
- Dynamic refinement removes completed subtasks, rechecks unresolved dependencies, and can introduce new subtasks based on execution progress.
- Algorithm 1 iterates over executable subtasks, updates completed and pending sets, refines the plan periodically, and returns the final reasoning state.
4 Data Construction
The data construction pipeline addresses limited generalization and inefficient traces by synthesizing dense, multi-type tasks from diverse corpora, graphs, subgraphs, and verified questions.
- Existing pipelines often emphasize deterministic entity-centric tasks and scale difficulty through reasoning hops, producing redundant evidence and inefficient traces.
- The proposed pipeline integrates corpus collection, graph construction, subgraph extraction, and verified question generation for deterministic and open-ended tasks.
- The raw corpus uses diverse real-world URLs spanning domains including art, sports, history, government, economics, politics, and science.
- URLs connected by information-seeking paths induce multi-hop relationships across documents and support graph-based task construction.
- Controlled random walks and breadth-first search extract task-specific subgraphs whose neighborhoods provide supporting evidence for target answers.
- Question synthesis encodes interdependencies as verifiable conditions and applies iterative verification or obfuscation to prevent premature inference.
- Open-ended research questions integrate evidence across subgraphs to elicit report-style explanation, comparison, and synthesis.
- Candidate trajectories undergo hard rejection and further quality filtering to enforce completeness, context validity, reasoning depth, tool usage, and format consistency.
5 Training Recipe
SMTL trains search agents on both deterministic and open-ended tasks using curated trajectories, supervised fine-tuning, and reinforcement learning. The recipe emphasizes efficient, information-rich interaction behavior.
- The SFT dataset combines Deep Search and Deep Research tasks built from a shared subgraph-based construction process.
- Deep Search: Deep Search tasks use 2-5-hop subgraphs and retain hierarchical question variants with ground-truth target entities.
- Deep Research: Deep Research tasks pose open-ended questions requiring broad exploration and synthesis across each subgraph.
- Trajectory Construction and Curation: Training trajectories are distilled from DeepSeek-V3.2 for Deep Search and GPT-5 for Deep Research.
- Trajectory Construction and Curation: Trajectory curation caps context at 64K tokens, requires at least 3 tool calls per step on average, and retains correct shortest trajectories.
- Reinforcement Learning: Reinforcement learning uses a modified RLOO procedure with outcome-based rewards for correct answers and zero reward for invalid tool-call formats.
6 Experiments
SMTL is evaluated across deterministic deep-search and open-ended deep-research benchmarks using LLM-based judging and multiple interaction settings. It combines competitive heterogeneous performance with shorter, parallelized search trajectories.
- Evaluation Setup: Evaluation spans BrowseComp, GAIA, XBench-DeepSearch, WebWalker-QA, FRAMES, SEAL-0, DeepResearch Bench RACE, and DeepResearch Gym.
- Evaluation Setup: Deep-search tasks use pass@1, while DeepResearch Bench RACE evaluates comprehensiveness, insight/depth, instruction following, and readability.
- Main Results: SMTL-100 reaches 43.6% on BrowseComp, 78.0% on XBench-DeepSearch, and 74.9% on WebWalker-QA, while SMTL-300 raises BrowseComp accuracy to 48.6%.
- Main Results: SMTL-100 scores 45.9% overall on DeepResearch Bench RACE, outperforming several open-source baselines and slightly exceeding Tongyi-DeepResearch-30B at 45.7%.
- Generalization: SMTL transfers the same parallel search framework from accuracy-driven benchmarks to long-form research synthesis without task-specific modification.
- Efficiency Analysis: SMTL localizes a key entity within 8 assistant turns, compared with 16 turns for MiroThinker-v1.0 in a BrowseComp case study.
- Efficiency Analysis: Parallel subtask exploration and staged re-planning, rather than deeper per-step reasoning, account for the case study’s shorter search trajectory.
7 Analysis
Analysis attributes SMTL’s efficiency to denser parallel evidence acquisition and shows that additional interaction budget and retrieval breadth mainly help difficult, evidence-limited cases. These changes preserve a favorable accuracy–complexity trade-off.
- Efficiency Evaluation: SMTL averages 3.5 tool calls per step, aggregating concurrent evidence and reducing redundant query reformulation.
- Efficiency Evaluation: SMTL-300 reaches 48.6% accuracy with 150.7 average steps and remains on the BrowseComp accuracy–step Pareto frontier.
- Maximum Interaction Steps: Successful trajectories do not show a noticeable increase in median steps as the budget grows, whereas failed trajectories usually terminate at the maximum budget.
- Maximum Interaction Steps: Larger interaction budgets improve difficult-case success by enabling alternative evidence-chain exploration and correction of suboptimal search directions.
- Retrieval Top-k: Increasing retrieval top-k from 4 to 8 raises SMTL-300 from 43.8 to 47.0 and SMTL-100 from 36.6 to above 41.8.
- Retrieval Top-k: Increasing top-k beyond 8 continues improving performance more slowly, indicating diminishing returns after informative candidates are included.
- Retrieval Top-k: The analysis identifies broader evidence acquisition as a more efficient scaling axis than simply extending reasoning length under fixed interaction budgets.
8 Conclusion
SMTL combines parallel evidence acquisition, structured context management, and unified data synthesis to improve efficiency and generalization in long-horizon search. Its results support search-centric scaling as a practical foundation for future deep research agents.
- SMTL replaces sequential reasoning with parallel evidence acquisition and integrates structured context management for efficient inference under constrained interaction budgets.
- Its automated data synthesis pipeline covers both deterministic question answering and open-ended research tasks.
- SMTL achieves state-of-the-art or competitive performance across diverse deep-search and deep-research benchmarks while reducing reasoning steps and inference latency.
- The results support prioritizing efficient, search-centric scaling over ever-deeper reasoning for future deep research agents.
9 Contributions
This section lists the authors and describes the minimal toolset used for web-scale information acquisition and consolidation.
- 9 Contributions: The author list includes Qianben Chen, Tianrui Qin, Jian Yang, Jiaheng Liu, and additional contributors.The supplied passages contain the author names but do not provide further contribution details.
- 9 Contributions: The agent uses two core tools: a search interface for candidate retrieval and a page-level crawler for extraction and goal-directed summarization.These tools support web-scale information acquisition and consolidation.
- 9 Contributions: web_search retrieves ranked results containing titles, snippets, and URLs to identify potentially relevant sources and guide subsequent crawling.Its default setting returns the top five entries through the Serper API interface to Google Search.
- 9 Contributions: crawl_page uses a target URL and explicit information goal to acquire page content and summarize material relevant to the current subtask.The page is crawled through the Jina Reader API and summarized by DeepSeek-V3.2.
B.1 Deep Search Data
The Deep Search data pipeline emphasizes objective, verifiable entity facts and constructs multi-hop questions without directly naming the entities involved.
- B.1 Deep Search Data: Deep Search extraction prioritizes precise numbers, specific times, and specific locations as objective, verifiable entity facts.Examples include founding years, employee counts, dates, countries, cities, distances, and elevations.
- B.1 Deep Search Data: The pipeline prohibits subjective, generic, vague, and unsupported functional descriptions.Information must be directly extracted, while unsupported pieces are ignored.
- B.1 Deep Search Data: Extracted information is organized into facts, a one-sentence relationship summary, and a multi-hop search question.The question is constructed strictly from extracted objective facts.
- B.1 Deep Search Data: Questions omit current and related entity names, instead using abstract or relational references such as “an organization” or “a company.”This preserves the need to identify the answer through search rather than direct name matching.
- B.1 Deep Search Data: The obfuscation stage selects one critical condition so the answer cannot be inferred directly and requires external search for verification.The description must retain only attributes explicitly present in the source text.
B.2 Deep Research Data
The Deep Research data pipeline creates open-ended, graph-grounded questions requiring multi-evidence synthesis and solves them through structured parallel tool use.
- B.2 Deep Research Data: Deep Research questions require integrating evidence across a supporting subgraph and produce report-style answers involving explanation, comparison, or synthesis.They are designed to admit multiple valid answers rather than a single exact factual output.
- B.2 Deep Research Data: The questions are graph-grounded, non-deterministic, multi-evidence, and not trivial reformulations of factual descriptions.They must be answerable from the provided subgraph while requiring reasoning over multiple entities, attributes, or relations.
- B.2 Deep Research Data: Both Deep Search and Deep Research use a shared parallel framework with graph-structured evidence, multiple subtasks, explicit planning, iterative refinement, and progress tracking.The design targets efficient long-horizon search under constrained interaction budgets.
- B.2 Deep Research Data: Deep Search uses a compact plan–plan-refine–answer structure, whereas Deep Research uses finer-grained subtasks and report-style synthesis.The two prompts differ in response format and control-flow granularity but share the same parallel execution paradigm.
- B.2 Deep Research Data: The execution protocol advances independent goals in parallel, uses tools for evidence gathering and verification, and refines actions from returned observations.The permitted tool interfaces include web_search and crawl_page.