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 must balance performance with computational efficiency, while generalization across heterogeneous research settings remains challenging. SMTL combines parallel evidence acquisition with multi-type task synthesis and achieves strong benchmark performance while reducing reasoning steps and latency.
Problem
Method
SMTL combines parallel evidence acquisition and structured context management with automated task synthesis spanning deterministic and open-ended settings.
Source: §1 · p. 2Results
Across deep-search and deep-research benchmarks, SMTL achieves state-of-the-art or competitive performance and reduces reasoning steps and latency; BrowseComp accuracy rises from 43.6% at 100 steps to 48.6% at 300.
Source: §6.2 · p. 8, §8 · pp. 11–12Takeaways & Limitations
The results support efficient, search-centric scaling as a practical and generalizable foundation for future deep research agents.
Source: §8 · p. 12
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 efficiency and generalization in long-horizon agentic search by replacing sequential reasoning with parallel evidence acquisition and synthesizing tasks across deterministic and open-ended settings. The approach reports state-of-the-art benchmark performance alongside fewer reasoning steps and lower inference latency.
1 Introduction
- Longer reasoning trajectories can improve task performance but increase inference latency, leaving efficient long-horizon search an open problem.The introduction frames task design and training trajectories for efficient long-horizon reasoning as a key question.Source: §1 · p. 2
- Deterministic question-answering and open-ended research tasks use different objectives and evaluation criteria, making cross-setting generalization challenging.Deterministic tasks are mainly measured by accuracy, while open-ended research emphasizes coverage, coherence, and synthesis quality.Source: §1 · p. 2
- BrowseComp performance reaches 44.3, with reasoning steps reduced by 78% and inference latency improved by up to 2.6×; the method reports state-of-the-art results across multiple benchmarks.The approach is trained end-to-end with supervised fine-tuning and reinforcement learning.Source: §1 · p. 2
- Parallel task decomposition and concurrent tool execution replace linear sequential reasoning, with structured context management for long-horizon inference under constrained context budgets.The framework uses plan-driven context management to acquire evidence in parallel.Source: §1 · p. 2
- An automated pipeline synthesizes representative search tasks across deterministic question-answering and open-ended research settings to support generalization.It reduces redundant samples while supporting generalization across task types.Source: §1 · p. 2
2 Related Work
Prior search-agent work centers on orchestration and multi-step tool use, while synthetic-data pipelines generate tool-interactive tasks through graph-based or easy-to-hard approaches. The related work identifies limited information-efficient search and limited support for open-ended research as gaps.
Agent Frameworks and Systems
- Agentic systems use external orchestration, planner–researcher or planner–executor workflows, reusable templates, and multi-agent coordination for search-intensive tasks.Examples include commercial deep-research systems, WebWeaver and OAgents, and MiroFlow’s benchmark-oriented framework (Anthropic, 2025; Google, 2024; OpenAI, 2025; Perplexity Team, 2025; Li et al., 2025f; Zhu et al., 2025a; MiroMind AI Team, 2025).Source: §2 · pp. 2–3
- Despite their architectural diversity, existing workflows often scale performance through deeper sequential reasoning and longer interaction horizons, reducing information efficiency.The passage attributes this inefficiency to computation spent on prolonged model-side reasoning rather than effective external evidence acquisition.Source: §2 · p. 3
Synthetic Data Pipelines
- Synthetic pipelines use graph-based generation, easy-to-hard expansion, or task expansion along depth and width, but emphasize difficulty or context length over information-efficient search and verification.WebSailor samples questions or trajectories from knowledge-graph subgraphs; WebShaper, ASearcher, and WebExplorer expand seed questions; TaskCraft adds incremental validation (Li et al., 2025a; Gao et al., 2025; Liu et al., 2025c; Tao et al., 2025a; Shi et al., 2025). Existing pipelines also mainly support deterministic QA or constrained structures, with limited support for open-ended research requiring flexible aggregation and cross-source validation.Source: §2 · p. 3
3 Parallel Agentic Workflow
The workflow replaces a single sequential reasoning chain with parallel evidence acquisition coordinated by a task plan that is refined as execution progresses. It constructs subtasks before tool use, executes ready subtasks concurrently, and synchronizes updates into a final reasoning state.
Algorithmic Outline
- The workflow initializes a plan, executes multiple subtasks in parallel, and periodically synchronizes intermediate results through plan refinement before producing an answer.It is designed to support concurrent execution and structured coordination across subtasks.Source: §3 · pp. 3–4
Initial Plan Construction
- Before tool execution, the agent decomposes a composite search task into interrelated, partially independent subtasks with concrete information-seeking or verification objectives.The initial plan exposes parallelizable paths early to enable concurrent evidence acquisition and higher information density.Source: §3 · p. 3
Parallel Execution and Tool Coordination
- At each timestep, ready pending subtasks run concurrently using available tools or agent actions, and their observations are aggregated into a unified reasoning state.Reusable external tools, including web search and page crawling, support concurrent information acquisition and verification.Source: §3 · p. 4
Dynamic Plan Refinement
- Periodic plan refinement removes completed subtasks, rechecks unresolved dependencies, and may introduce new subtasks based on the current execution state.The algorithm performs refinement at intervals while parallel execution continues.Source: §3 · p. 4
4 Data Construction
SMTL addresses limits in existing data pipelines by synthesizing diverse, information-dense tasks for both deterministic question answering and open-ended research. Its pipeline builds graph-grounded evidence structures and verified questions to support multi-hop, cross-source reasoning.
4.1 Deep Search Data
- TaskCraft trajectories (Shi et al., 2025) supply diverse, relationally linked URLs, which are parsed with Jina and organized into a knowledge graph using LightRAG (Guo et al., 2024).The pipeline extracts entities and relationships from text chunks and uses embedding-plus-reranker retrieval to support graph construction.Source: §4.1 · p. 5
- Controlled random walks and BFS extract compact, coherent subgraphs, while hierarchical question synthesis turns their multi-hop evidence into questions with cross-validated paths and leakage checks.Hop depth and branching factor control difficulty; synthesis is verified iteratively, restructuring or obfuscating questions when answers are inferable prematurely, for up to five iterations.Source: §4.1 · pp. 5–6
4.2 Deep Research Data
- Open-ended research questions are generated from task subgraphs to elicit report-style answers that integrate evidence through explanation, comparison, and synthesis across sources.The questions are designed to encourage long-horizon planning, information aggregation, and cross-source reasoning.Source: §4.2 · p. 6
- Candidate research trajectories pass rule-based checks for structural validity and depth, then LLM-as-a-Judge assessment of semantic quality, with only trajectories passing both retained as reference outputs.The checks cover completeness, context validity, reasoning depth, tool use, format consistency, comprehensiveness, insight, instruction following, and readability.Source: §4.2 · p. 6
5 Training Recipe
The training recipe combines supervised fine-tuning on Deep Search and Deep Research tasks with reinforcement learning. It constructs and curates task-specific trajectories, then optimizes them using outcome-based rewards and modified RLOO updates.
5.1 Post-training Supervised Fine-tuning.
- SFT combines Deep Search questions with entity answers and open-ended Deep Research questions that encourage exploration and synthesis across subgraphs.Deep Search uses 2–5-hop subgraphs and retains hierarchical question variants while filtering overrepresented target entities.Source: §5.1 · pp. 6–7
- Deep Search trajectories are distilled from DeepSeek-V3.2 (Liu et al., 2025a), while Deep Research trajectories are distilled from GPT-5 (OpenAI, 2025).For tasks with multiple successful trajectories, curation retains only correct trajectories with the shortest interaction length.Source: §5.1 · p. 7
- Trajectory curation caps total length at 64K tokens and requires an average of at least 3 tool calls per step.These constraints limit redundant interactions and noisy supervision while encouraging active information acquisition.Source: §5.1 · p. 7
5.2 Reinforcement Learning.
- RLOO uses token-level loss, sequence-level importance sampling for rollout correction, and filters environmental failures and excessively long or maximum-turn trajectories.The modified REINFORCE Leave-One-Out method (Ahmadian et al., 2024) uses an unbiased advantage estimator versus GRPO; its token-level loss follows DAPO (Yu et al., 2025), and rollout correction addresses log-probability discrepancies.Source: §5.2 · p. 7
- RL uses outcome-based rewards: an LLM judge assigns 1 for a correct final answer and 0 otherwise, while malformed tool calls terminate generation with reward 0.This explicitly encourages correct tool usage.Source: §5.2 · p. 7
6 Experiments
SMTL performs strongly across deep-search and open-ended research benchmarks. Its efficiency comes from parallel subtask exploration, staged replanning, and plan-based context management.
6.1 Experimental Setup
- The evaluation spans six deep-search benchmarks and DeepResearch Bench RACE, using pass@1 for deep search and judge-based scores for long-form research reports.RACE reports an overall score and Comprehensiveness, Insight/Depth, Instruction Following, and Readability criteria.Source: §6.1 · pp. 7–8
- When history reaches the 128K context budget without a confirmed answer, SMTL forces plan refinement, drops pre-plan context, and continues from the refreshed plan.The agent otherwise refines its plan every N=5 steps; SMTL-100 and SMTL-300 denote maximum interaction limits of 100 and 300 steps, while DeepResearch Bench uses the 100-step setting.Source: §6.1 · p. 8
6.2 Main Results
- BrowseComp accuracy rose from 43.6% (SMTL-100) to 48.6% (SMTL-300), while SMTL-100 exceeded Tongyi-DeepResearch-30B (43.4%) and MiroThinker-v1.0-30B (41.2%).Across BrowseComp budgets from 50 to 300 steps, SMTL remained on the accuracy-versus-trajectory-length Pareto frontier.Source: §6.2 · p. 8, Table 1 · p. 9
- SMTL-100 scored 45.9% overall on DeepResearch Bench RACE, slightly surpassing Tongyi-DeepResearch-30B (45.7%) and Kimi-Researcher (44.6%).Its scores were 42.1% for Comprehensiveness, 45.6% for Insight/Depth, 49.6% for Instruction Following, and 45.5% for Readability; it transferred to open-ended research without task-specific modification.Source: §6.2 · pp. 8–9, Table 1 · p. 9
- On BrowseComp, SMTL localized the key entity in 8 assistant turns, compared with 16 turns for MiroThinker-v1.0-30B.SMTL explores hypothesized subtasks in parallel and replans from intermediate observations, whereas MiroThinker-v1.0 permits one tool call per round and proceeds sequentially.Source: §6.2 · pp. 9–10
7 Analysis
SMTL balances accuracy with fewer interaction steps through parallel evidence acquisition, while broader retrieval and larger interaction budgets improve performance by helping the agent explore difficult cases. The analyses indicate that these gains come from denser evidence gathering and continued search, rather than simply extending reasoning depth.
7.1 Efficiency Evaluation
- SMTL-100 reaches 44.6% accuracy in 60.4 assistant steps, versus Tongyi-DeepResearch-30B’s 43.4% in 75.2 steps and MiroThinker-v1.0-30B’s 41.2% in 206.0 steps; SMTL-300 reaches 48.6% in 150.7 steps.SMTL remains on the accuracy–step Pareto frontier across interaction budgets, according to the analysis.Source: §7.1 · p. 10, Table 2 · p. 10
- SMTL averages 3.5 tool calls per step, acquiring evidence concurrently across subtasks and reducing redundant query reformulation.Aggregating more information within each interaction round increases information density per step and enables shorter, more effective trajectories.Source: §7.1 · p. 10
7.2 Ablations on Maximum Interaction Steps.
- Successful trajectories generally finish before the interaction limit, while failed trajectories often reach it, so higher budgets mainly increase steps spent on failed cases.The analysis attributes many failures to exhausting the interaction budget rather than premature incorrect answers, and interprets continued search as exploring alternatives rather than misunderstanding or overconfidence.Source: §7.2 · p. 10
- Larger interaction budgets improve success on difficult cases by allowing more evidence-chain exploration and periodic plan refinement to redirect search toward promising subtasks.The gains primarily alleviate search-budget constraints rather than compensate for systematic reasoning errors.Source: §7.2 · pp. 10–11
7.3 Ablations on Retrieval Top-k.
- Increasing 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; gains continue more slowly from 8 to 20.The results indicate diminishing returns as the most informative candidates enter the retrieval window, while additional results still reduce the risk of missing critical evidence.Source: §7.3 · p. 11
- At a fixed interaction budget, broader retrieval packs more candidate evidence into each search action, making retrieval breadth a more efficient scaling axis than longer reasoning.The analysis frames this as consistent with SMTL’s design intuition that expanding evidence acquisition can improve long-horizon search without extending reasoning depth.Source: §7.3 · p. 11
8 Conclusion
SMTL combines parallel evidence acquisition, structured context management, and automated data synthesis to support efficient long-horizon search across deterministic and open-ended research tasks. The paper reports competitive benchmark performance with fewer reasoning steps and lower inference latency.
8 Conclusion
- SMTL replaces sequential reasoning with parallel evidence acquisition for efficient long-horizon inference under constrained interaction budgets.Its framework integrates an efficient agentic workflow with structured context management.Source: §8 · p. 11
- SMTL’s automated data synthesis pipeline supports both deterministic question answering and open-ended research tasks.The agent is trained end-to-end with supervised fine-tuning and reinforcement learning.Source: §8 · p. 11
- SMTL achieves state-of-the-art or competitive performance across deep search and deep research benchmarks while substantially reducing reasoning steps and inference latency.The conclusion presents these results as evidence for efficient, search-centric scaling rather than ever-deeper reasoning.Source: §8 · pp. 11–12
- The conclusion presents efficient, search-centric scaling as a practical and generalizable foundation for future deep research agents.The authors hope the work encourages further exploration of efficiency-oriented agentic designs.Source: §8 · p. 12
A Tools Setup
The agent uses a minimal two-tool setup for web-scale information acquisition and consolidation: search retrieves candidate sources, and a page crawler extracts and summarizes content.
A Tools Setup
- web_search uses Serper to return ranked Google results, with the default query retrieving five titles, snippets, and URLs to guide crawling.These results provide high-level signals for identifying potentially relevant sources.Source: A Tools Setup · p. 17
- crawl_page uses Jina Reader and DeepSeek-V3.2 (Liu et al., 2025a) to summarize a URL’s content against an explicit information goal.Goal-conditioned summarization targets evidence relevant to the current subtask and reduces irrelevant context propagation.Source: A Tools Setup · p. 17
B Data Construction · B.1 Deep Search Data
B.1 Deep Search Data describes prompts for extracting, evaluating, generating, obfuscating, and verifying entity-centered information. The pipeline emphasizes explicit, verifiable facts and constructs search questions that require identifying entities through their relations and attributes.
B.1 Deep Search Data
- The extraction and description-evaluation prompts require explicit textual evidence, score factuality from 0 to 100, and copy key factual phrases verbatim.The rubric defines High as 80–100, Medium as 50–79, and Low as 0–49; missing extraction fields may be returned empty rather than inferred or fabricated.Source: §B.1 · pp. 17–18
- The generation prompt prioritizes precise numbers, time, locations, events, and measurable attributes, organizing objective facts hierarchically into facts, a relationship summary, and a multi-hop question.It prohibits subjective or generic descriptions, requires factual details to be directly extracted, and directs that questions omit entity names and include the extracted objective information.Source: §B.1 · pp. 19–20
- The fact-extraction procedure categorizes verifiable details as numerical, temporal, locational, event-based, or measurement-based and requires each fact to contain an objective element.The question-construction prompt uses only extracted facts, includes their objective details, and returns an empty result when there is insufficient evidence.Source: §B.1 · p. 20
- The obfuscation prompt selects a critical condition to obscure so the answer cannot be inferred directly but remains discoverable through search.Strategies include making time or quantities less precise, replacing names with searchable attributes, deleting revealing details, abstracting concepts, or combining lightweight changes while preserving coherence.Source: §B.1 · pp. 20–21
- Answer verification identifies the most specific entity supported by a description and returns “Cannot determine” when the description is insufficient.The response must name an entity rather than a concept or category and rely only on the supplied description.Source: §B.1 · pp. 21–22
- Answer-equivalence verification accepts different phrasings, aliases, abbreviations, and alternative names when they refer to the same entity.It returns false for “Cannot determine” or a clearly incorrect answer, and rejects concepts, categories, and abstract descriptions as predictions.Source: §B.1 · p. 22
B.2 Deep Research Data · C Case Study · D Prompts
This section details the appendix’s deep-research task and parallel-search design, including a case study contrasting SMTL with MiroThinker. It also specifies the LLM-as-a-Judge output format and parsing requirements.
B.2 Deep Research Data
- Open-ended research questions must synthesize multiple graph entities and relations, admit multiple supported answers, and avoid single-fact or short-answer formulations.The prompt requires questions to be answerable from the supplied subgraph and suitable for report-style explanation, comparison, or synthesis.Source: §B.2 · p. 23
C Case Study
- SMTL’s parallel subtasks and staged re-planning enable faster evidence localization and verification than MiroThinker-v1.0-30B’s strictly sequential search.The case study illustration compares SMTL-30B and MiroThinker-v1.0-30B.Source: Figure 5 · p. 24
D.1 System Prompts
- Both system prompts use graph-structured evidence, goal or subtask decomposition, parallel execution, coordinated tools, explicit planning, iterative refinement, and progress tracking.Deep Search uses a compact plan–plan-refine–answer structure for deterministic questions; Deep Research uses finer-grained subtasks and report-style synthesis for open-ended tasks.Source: §D.1 · pp. 24–25
D.2 Summary Prompt
- The webpage-summary prompt scans for goal-relevant content, extracts supporting evidence while preserving context, and organizes it into a concise, coherent summary.Source: §D.2 · p. 27
D.3 LLM-as-a-Judge Prompts
- The two judge prompts match their task outputs: Deep Search checks semantic equivalence to ground truth, while Deep Research scores comprehensiveness, insight, instruction following, and readability.Deep Research comparisons assign each article a separate 0–10 score for each criterion.Source: §D.3 · pp. 28–29
- The prompt defines article_1_score and article_2_score as 0–10 fields and requires fully parsable JSON with parsing-sensitive characters escaped.The requirements appear in the LLM-as-a-Judge prompt.Source: §D.3 · p. 30