PaperlayerView exact revision on arXiv

Source-linked AI summary

WideSeek: Advancing Wide Research via Multi-Agent Scaling

Ziyang Huang, Haolin Ren, Xiaowei Yuan, Jiawei Wang, Zhongtao Jiang, Kun Xu, Shizhu He, Jun Zhao, Kang Liu

arXiv:2602.02636v1cs.CLcs.AIcs.IR

TL;DR

Wide Research lacks benchmarks and optimization methods for retrieving broad, constraint-satisfying information in parallel. The paper introduces WideSeekBench and WideSeek, then shows that dynamic multi-agent scaling with end-to-end RL improves broad-search performance while transferring to deep-research tasks.

  • Problem

    Wide Research lacks dedicated benchmarks and optimization methodologies for retrieving and synthesizing large-scale information under complex constraints.

  • Method

    The paper constructs WideSeekBench through a multi-phase knowledge-graph pipeline and introduces WideSeek, a dynamic hierarchical multi-agent architecture trained by linearizing shared multi-agent trajectories for end-to-end RL.

  • Results

    WideSeek’s dynamic orchestration and end-to-end multi-agent optimization advance broad parallel retrieval, with WideSeek-8B-RL reaching 26.42% accuracy on BrowseComp-Plus, +12.20% over the base model.

  • Takeaways & Limitations

    Scaling sub-agents and tool use is a supported direction for improving Wide Research capabilities and can transfer from wide-research training to deep-research tasks.

Abstract

from arXiv · show

Search intelligence is evolving from Deep Research to Wide Research, a paradigm essential for retrieving and synthesizing comprehensive information under complex constraints in parallel. However, progress in this field is impeded by the lack of dedicated benchmarks and optimization methodologies for search breadth. To address these challenges, we take a deep dive into Wide Research from two perspectives: Data Pipeline and Agent Optimization. First, we produce WideSeekBench, a General Broad Information Seeking (GBIS) benchmark constructed via a rigorous multi-phase data pipeline to ensure diversity across the target information volume, logical constraints, and domains. Second, we introduce WideSeek, a dynamic hierarchical multi-agent architecture that can autonomously fork parallel sub-agents based on task requirements. Furthermore, we design a unified training framework that linearizes multi-agent trajectories and optimizes the system using end-to-end RL. Experimental results demonstrate the effectiveness of WideSeek and multi-agent RL, highlighting that scaling the number of agents is a promising direction for advancing the Wide Research paradigm.

1. Introduction

Wide Research extends search from deep, sequential retrieval toward broad, parallel synthesis of information satisfying complex constraints. The paper addresses benchmark and agent-design gaps by introducing WideSeekBench and a dynamic multi-agent system optimized with end-to-end reinforcement learning.

  • Motivation: Wide Research replaces deep-but-narrow sequential reasoning with high-breadth parallel orchestration across expansive search spaces.Its agents decompose global objectives into parallel subtasks and can iteratively research and cross-validate results.
  • Motivation: Wide Research targets comprehensive collections of information satisfying complex constraints, such as comparative tables assembled from many sources.This setting is framed as General Broad Information Seeking and is tied to real-world productivity scenarios.
  • Challenges: Existing benchmarks are limited in scale, diversity, categorization depth, and training-data availability, while synthesis methods mainly optimize search depth.These limitations hinder evaluation and optimization for broad information retrieval under complex constraints.
  • Contributions: WideSeekBench provides a benchmark for General Broad Information Seeking, constructed through a rigorous pipeline designed for broad-search evaluation.The benchmark is presented as the paper’s response to the lack of dedicated resources for Wide Research.
  • Agent Optimization: WideSeek dynamically instantiates any number of sub-agents as needed, enabling parallel and serial execution under a Planner-Executor architecture.The main agent plans, decomposes tasks, and self-reflects, while sub-agents execute subtasks and tool calls.
  • Results: Experiments support shifting agent design from sequential reasoning to dynamic parallel orchestration and specialized end-to-end multi-agent optimization.The conclusion identifies agent scaling as a route toward searching at scale in complex scenarios.

2. Data Pipeline & Benchmark

WideSeekBench formalizes broad information seeking as constructing structured tables from entities satisfying complex semantic constraints. Its multi-phase pipeline uses knowledge-graph sampling, logical composition, schema extension, LLM-assisted task synthesis, filtering, and balanced evaluation.

  • Task Definition: GBIS retrieves entities satisfying a complex semantic constraint and synthesizes their required attributes into a structured ground-truth table.A task consists of a query Q and attribute set A; the query induces a filter Φ that determines the target entity set E∗.
  • Seed Constraint Construction: The benchmark pipeline starts with human-defined domains, sampled sub-domains, seed entities, and atomic constraints extracted from a knowledge graph.This top-down strategy is intended to provide coverage and diversity across domains.
  • Logical Composition & Schema Extension: Atomic constraints are recursively composed with conjunction, disjunction, and negation, then executed over the knowledge graph to obtain target entities and extend their attribute schema.Selected attributes must provide entity coverage and sufficient value diversity before populating the target table.
  • Agent Task Synthesis: LLM-based task synthesis self-refines queries by checking reconstructed logic against the intended constraint and generates column-wise rubrics from attribute semantics and values.This phase produces approximately 15,000 candidate tasks.
  • Multi-Stage Filtering: Multi-stage filtering combines web grounding and sparsity rules, LLM scoring on five quality dimensions, and human verification, yielding 5156 final tasks.The filters assess human-likeness, solvability, common sense, temporal stability, and rubric rationality.
  • Benchmark: Test tasks are classified by target-information volume using the total number of ground-truth table cells, with tasks divided into 10 intervals.The benchmark also provides a simulated environment with a local corpus and search engine for reproducible evaluation.

3. WideSeek

WideSeek addresses broad retrieval tasks with a dynamic hierarchical multi-agent system that decomposes work into parallel subtasks and unifies the resulting trajectories for end-to-end reinforcement learning. Its policy dynamically determines sub-agent creation, while unified rewards optimize retrieval correctness and tool-use format.

  • 3. WideSeek: WideSeek constructs a structured answer table satisfying a complex semantic constraint through a dynamic hierarchical multi-agent system.The system uses a unified policy to retrieve entities and attribute values for the target table.
  • 3.1. Multi-Agent Rollout: The Main Agent plans and terminates rollouts, dynamically choosing when and how many Sub-Agents to fork in parallel.The number of Sub-Agents is determined by the policy rather than fixed as a hyperparameter.
  • 3.1. Multi-Agent Rollout: Each Sub-Agent executes its local sub-task, returns a textual sub-result, and updates the global state used for final synthesis.The composite trajectory interleaves Main Agent reasoning with execution traces from dynamically created Sub-Agents.
  • 3.3. Multi-Agent Reinforcement Learning: Unified multi-agent RL optimizes the dynamic execution tree as a single generative process using GRPO over groups of unified trajectories.The global reward combines Item-F1 correctness with a format-violation penalty, and GRPO uses rewards relative to the sampled group.
  • 3.3. Multi-Agent Reinforcement Learning: The system linearizes hierarchical multi-agent traces into one unified trajectory because all agents share the same policy checkpoint.The unified trajectory interleaves Main Agent steps with the trajectories of Sub-Agents forked at those steps.

4. Experiment

Experiments evaluate WideSeek on WideSeekBench and test generalization to BrowseComp-Plus. Specialized optimization substantially increases search effort and performance, while larger-scale tasks remain difficult.

  • Experimental Setup: WideSeekBench evaluates proprietary and open-source models, while BrowseComp-Plus tests whether WideSeek generalizes to Deep Research tasks.The experiments use Qwen3-8B as the base model for agent optimization.
  • Main Results: Mean@4 Item-F1 reaches only 21.03 for current state-of-the-art proprietary models on WideSeekBench, highlighting the difficulty of broad search.DeepSeek-v3.2 forks 31.25 sub-agents, while GPT-5.2 executes 408 tool calls, yet coordination remains insufficient for complex constraints.
  • Main Results: WideSeek-8B-SFT-RL achieves Item F1 of 12.87% and Max Row F1 of 3.88%, with tool calls increasing 28.82× over base.The same model increases sub-agent instantiation by 6.36×; RL from scratch also learns to scale search actions.
  • Generalization: On BrowseComp-Plus, WideSeek with base Qwen3-8B scores 14.22%, while WideSeek-8B-RL reaches 26.42%, a +12.20% improvement over base.The scaffold also outperforms Qwen3-32B using ReAct, which scores 10.72%.

5. Analysis

Analysis across information volume, logical constraints, and domains shows consistent benefits from multi-agent RL, alongside clear limits on massive-scale retrieval and set-difference reasoning.

  • Overall Analysis: Multi-agent RL consistently enhances performance across target-information volume, constraint type, and domain dimensions.WideSeekBench supports this granular evaluation through multi-dimensional task classification.
  • Volume of Target Information: Across all target-information intervals, the performance hierarchy is WideSeek-8B-SFT-RL > WideSeek-8B-SFT > WideSeek-8B-RL.Performance gaps are minimal for volumes from [4, 128].
  • Volume of Target Information: Performance significantly degrades as target-information volume rises from [128, 4096], making massive-scale information seeking a formidable challenge.In [2048, 4096], SFT-based models show reduced tool-call frequency alongside low success rates, consistent with early stopping.
  • Constraint Type: OR constraints generally achieve higher performance, whereas NOT constraints yield the lowest performance among the seven logical constraint types.The analysis identifies set difference operations as a distinct reasoning bottleneck.
  • Domain: Across 18 domains, the hierarchy WideSeek-8B-SFT-RL > WideSeek-8B-SFT > WideSeek-8B-RL holds universally, despite domain sensitivity.Infrastructure is cited as one domain with notably higher performance.

6. Related Work

Related work has largely optimized search depth or used static multi-agent coordination. WideSeek instead formalizes search width with knowledge graphs and jointly optimizes orchestration end to end.

  • Data Synthesis for Search Agent: Earlier synthetic-data methods simulate web navigation paths, information aggregation, or entity-coreference networks, primarily targeting search depth.Examples include WebWalkerQA, WebDancer, and WebSailor.
  • Data Synthesis for Search Agent: WideSeek uses a set-theoretic formalization and Knowledge Graph clusters to regulate task breadth and constraint complexity.Set operators define target entity sets within expansive search spaces.
  • Multi-Agent Optimization: Traditional LLM multi-agent systems commonly use static, heuristic-driven roles and lack parameter-level optimization for specific collaborative tasks.Recent branch-return methods often fork a fixed number of sub-agents at each step.
  • Multi-Agent Optimization: Unlike orchestration-only optimization, WideSeek applies end-to-end reinforcement learning to optimize planner–executor interaction jointly.This targets interaction granularity between the planner and executors.

7. Conclusion

The paper introduces a benchmark and an end-to-end optimized dynamic multi-agent system for Wide Research. Its experiments support agent scaling for complex parallel retrieval tasks.

  • Contributions: WideSeekBench formalizes General Broad Information Seeking through a multi-phase pipeline that mines intersected world knowledge from Knowledge Graphs.The benchmark addresses the shift from Deep to Wide Research.
  • Contributions: WideSeek is a dynamic hierarchical multi-agent architecture optimized through end-to-end reinforcement learning.The system is designed to leverage agent scaling for complex, parallel retrieval tasks.
  • Conclusion: The authors conclude that specialized multi-agent optimization enables models to search at scale in complex scenarios.The conclusion frames the transition from sequential to dynamic parallel orchestration as a fundamental design shift.

A.1. Benchmark Comparison

WideSeekBench is positioned as a broad, structured information-seeking benchmark, with a pipeline built around Wikidata entities, domain coverage, and information-density ranking.

  • Benchmark Comparison: WideSeekBench compares broad structured information collection with benchmarks targeting specific hidden information.The comparison distinguishes task type, automated generation, and fine-grained multidimensional classification.
  • Data Pipeline: The pipeline uses Wikidata entities and relations as the basis for constructing benchmark tasks.A local QLever engine supports complex SPARQL queries over the full knowledge graph, while seed entities provide the semantic basis for synthesis.
  • Domain Coverage: The taxonomy defines 18 high-level domains mapped to 200 Wikidata-based sub-domains.These sub-domains establish a controlled search scope for benchmark construction.
  • Seed Selection: For each sub-domain, the pipeline identifies 80 informative seed entities through recursive retrieval and information-density ranking.Candidates are retrieved through subclass closure and ranked by outgoing RDF triple counts.

A.4. Logical Composition and Task Synthesis

The pipeline composes atomic knowledge-graph constraints into validated retrievals and bounded tables, controlling both logical complexity and table size before populating cell values.

  • Logical Composition: Each seed entity contributes up to 200 composite constraints, with at most 4 validated tables retained to control redundancy and balance.The cap limits the number of retained tables per seed after validation.
  • Logical Composition: Atomic constraints are composed into composite SPARQL filters using seven logical patterns, with 1–8 atomic constraints per composite constraint.The domain constraint is excluded from this atomic-constraint count.
  • Execution and Verification: Candidate entity sets are retained only when their cardinality falls within [1, 1024].Queries outside this interval are discarded before attribute retrieval.
  • Table Construction: Target attributes require at least 50% entity coverage and sufficient value diversity, while table sizes must contain 8–8192 cells.Invalid labels are removed before batch retrieval populates the final table.

A.5. Agent Task Synthesis and Multi-Stage Filtering

WideSeekBench converts structured filters into human-like search tasks through randomized language generation, adaptive rubric synthesis, and layered quality assurance.

  • Task Synthesis: A cyclic generation-verification pipeline transforms logical filters into diverse, human-like search tasks.LLM-based operations in this subsection use GPT-5.
  • Task Synthesis: A generator model rewrites structured templates into natural language while randomizing among ten syntactic modes.The process is designed to diversify phrasing while preserving logical content.
  • Rubric Synthesis: Adaptive rubrics accept aliases, semantic date equivalence, numeric tolerances, and order-independent set equality according to data type.These criteria replace rigid string matching with distribution-aware compliance standards.
  • Quality Assurance: Quality assurance combines rule-based filtering, five-category LLM judging, and human verification.Tasks are rejected for failures involving human-likeness, solvability, common sense, temporal or wiki fit, or rubric rationality.

A.6. WideSeekBench Statistics

WideSeekBench contains thousands of tasks spanning broad domains and varying logical complexity, with statistics designed to expose differences in information volume, constraints, and topical coverage.

  • Scale: 4,436 training tasks and 720 test tasks cover 18 domains.The benchmark statistics report the dataset scale across training and test splits.
  • Constraint Complexity: Training constraints are dominated by AND patterns at 37.8%, while AND NOT accounts for 19.5%.The test set reduces simple AND patterns to 20.0% and increases complex composites.
  • Domain Diversity: WideSeekBench spans dominant media and gaming topics as well as scientific, technical, life-science, and natural-geography concepts.Examples include film, video games, airports, protein families, enzymes, planetary nebulae, and glaciers.
  • Figures: Figure 7 organizes scale statistics by training and test sets, while Figure 8 uses word clouds whose term sizes indicate frequency.The figures present dataset scale and topical diversity across the two splits.

A.7. Task Cases

The task cases span multiple domains and query structures, requiring systems to retrieve entities under conjunctive, disjunctive, and exclusion constraints and return specified attributes in tables.

  • Scientific journal: Scientific-journal tasks request journals meeting Wiley, geography, and English-related conditions, with inception, origin, indexing, and Danish indicator attributes.
  • Evaluation criteria: Column rubrics accept semantically equivalent values, including recognized aliases, format variations, numerical equivalence, and order-independent sets.
  • Lake: Lake tasks retrieve entities under geographic, basin-country, and exclusion conditions, returning elevation, coordinates, and country.
  • Television series: Television-series tasks combine alternative logical patterns, including island-based science fiction and relationships involving Ashes to Ashes and its spin-off.

A.9. Evaluation

The evaluation measures generated-table quality at table, row, and cell granularities using an LLM judge and column-wise rubrics.

  • The framework uses GPT-4.1 as the default judge LLM with column-wise rubrics for cell alignment.
  • Success Rate requires an answer table to match the ground truth exactly in both content and structure.
  • Row F1 measures retrieval and generation accuracy by matching rows only when every cell agrees with the corresponding ground-truth row.
  • Item F1 evaluates individual-cell correctness, regardless of whether the complete row is correct.

B. Experiments

The cold-start procedure builds a supervised policy from teacher-generated trajectories filtered by Item F1, then optimizes it with token-level negative log-likelihood.

  • Multiple teacher policies generate diverse rollout trajectories for each training query.
  • Trajectories enter the SFT dataset only when their Item-level F1 exceeds the filtering threshold η.
  • The filtering threshold is η = 0.6.
  • SFT initializes the policy by minimizing standard negative log-likelihood over the filtered high-quality trajectories.
  • The loss is computed only on tokens generated by the model, including thoughts and actions.

B.2. Training Dynamics

WideSeek training uses function calling to create parallel sub-agents, with training dynamics tracking rewards and tool calls over time.

  • The training-dynamics figure tracks reward evolution and tool-call times throughout the entire training process.
  • The implementation uses function calling to create sub-agents for the main agent.
  • The create_sub_agent tool accepts a variable-sized list of tasks for parallel information gathering.
  • Table 9 reports the hyperparameters used for RL training.
  • Each sub-agent task includes a unique agent identifier and a specific task description.

B.4. Case Study

The case study illustrates WideSeekBench’s broad information-seeking setup through multi-agent trajectories and film-research tasks combining complex constraints with many requested attributes.

  • Trajectory comparison: Figure 10 compares unified task trajectories from Qwen3-30B-A3B-Thinking, WideSeek-8B-SFT-RL, WideSeek-8B-SFT, and WideSeek-8B-RL.The figure also includes a case trajectory for WideSeek-8B-RL.
  • Query constraints: The case query intersects Academy Award for Best Sound nominations with National Board of Review Top Ten Films, while excluding films linked to two named people.The requested output contains numerous film attributes, including color, producer, review score, ratings, locations, distribution, and production details.
  • Output schema: The benchmark’s evaluation schema specifies typed fields such as film, color, country of origin, producer, review score, and multiple rating categories.The field specifications define formats for values including color, country, producer, review score, CNC rating, assessment, narrative location, and distributor.
  • Retrieved candidates: The reported intersection includes films such as The Love Parade, City Lights, I Am a Fugitive from a Chain Gang, Topaze, It Happened One Night, and The Informer.The listed films are presented as both National Board selections and Academy Award nominees for their respective years.
  • Retrieved attributes: The case outputs provide detailed attributes for films including It Happened One Night, The Life of Emile Zola, and The Best Years of Our Lives.Examples include director, duration, country of origin, producer, review score, and narrative or production information.
  • Evidence gaps: The case study records missing source information for several ratings and production-designer fields.For A Star Is Born, CNC, Kijkwijzer, FSK, and production-designer information are reported as unavailable or unspecified in the accessible sources.
Loading 2602.02636v1…