PaperlayerView exact revision on arXiv

Source-linked AI summary

WebLeaper: Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking

Zhengwei Tao, Haiyang Shen, Baixuan Li, Wenbiao Yin, Jialong Wu, Kuan Li, Zhongwang Zhang, Huifeng Yin, Rui Ye, Liwen Zhang, Xinyu Wang, Pengjun Xie, Jingren Zhou, Yong Jiang

arXiv:2510.24697v1cs.CL

TL;DR

WebLeaper targets inefficient information seeking, which the paper links in part to sparse target entities in training tasks. It constructs tree-structured, entity-rich tasks and filters trajectories for accuracy and efficiency, reporting consistent improvements across five benchmarks.

  • Problem

    Prior information-seeking work emphasizes search depth, while agents often search inefficiently; sparse target entities in training tasks limit opportunities to learn efficient behaviors.

    Source: §1 · p. 2

  • Method

    WebLeaper synthesizes tree-structured Basic, Union, and Reverse-Union tasks and curates trajectories with ISR and ISE for accuracy and efficiency.

    Source: §1 · p. 2, §3.2 · p. 7

  • Results

    WebLeaper achieves consistent improvements over strong baselines across five public information-seeking benchmarks.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The paper identifies sparse target entities in conventional training tasks as a primary contributor to inefficient information seeking.

    Source: §6 · p. 14

    The ISE stability result assumes entity-discovery step counts are i.i.d. with finite positive mean.

    Source: §2.3 · p. 4

Abstract

from arXiv · show

Large Language Model (LLM)-based agents have emerged as a transformative approach for open-ended problem solving, with information seeking (IS) being a core capability that enables autonomous reasoning and decision-making. While prior research has largely focused on improving retrieval depth, we observe that current IS agents often suffer from low search efficiency, which in turn constrains overall performance. A key factor underlying this inefficiency is the sparsity of target entities in training tasks, which limits opportunities for agents to learn and generalize efficient search behaviors. To address these challenges, we propose WebLeaper, a framework for constructing high-coverage IS tasks and generating efficient solution trajectories. We formulate IS as a tree-structured reasoning problem, enabling a substantially larger set of target entities to be embedded within a constrained context. Leveraging curated Wikipedia tables, we propose three variants for synthesizing IS tasks, Basic, Union, and Reverse-Union, to systematically increase both IS efficiency and efficacy. Finally, we curate training trajectories by retaining only those that are simultaneously accurate and efficient, ensuring that the model is optimized for both correctness and search performance. Extensive experiments on both basic and comprehensive settings, conducted on five IS benchmarks, BrowserComp, GAIA, xbench-DeepSearch, WideSearch, and Seal-0, demonstrate that our method consistently achieves improvements in both effectiveness and efficiency over strong baselines.

1 Introduction

The introduction identifies inefficient information seeking and sparse target entities in training tasks as key challenges. WebLeaper addresses them with entity-rich task variants and efficient, accurate trajectories, and reports consistent improvements across five benchmarks.

1 Introduction

  • Current information-seeking agents often search inefficiently, with a competitive agent’s valid-action distribution peaking around 0.04.The paper associates inefficient behavior with redundant query reformulations, irrelevant retrievals, and unnecessarily long search chains.Source: §1 · p. 2, Figure 2 · p. 2
  • Sparse target entities in conventional training tasks limit exposure to informative cues, leading agents to process irrelevant content and weakening search strategies.The paper also says this sparsity can bias efficiency measurement and impede learning efficient search behaviors.Source: §1 · p. 2
  • WebLeaper models information seeking as tree-structured reasoning and synthesizes Basic, Union, and Reverse-Union tasks, then retains trajectories that are accurate and efficient.The tasks use curated Wikipedia tables; ISR and ISE guide trajectory filtering and are incorporated into a hybrid reinforcement-learning reward.Source: §1 · pp. 2–3
  • WebLeaper achieves consistent improvements over strong baselines across five public information-seeking benchmarks.The benchmarks are BrowserComp, GAIA, xbench-DeepSearch, WideSearch, and Seal-0.Source: §1 · p. 3

2 Definitions

The definitions section treats information seeking as collecting required entities through a sequential agent process. It introduces measures of information coverage and efficiency, and argues that ISE becomes more stable as target-entity counts grow.

2.1 Information-Seeking Task

  • An information-seeking task collects the set of target entities required by a question, including intermediate entities needed along the reasoning path.The required set is not limited to the final explicit answers.Source: §2.1 · p. 3

2.2 Information-Seeking Agent

  • The agent proceeds through discrete steps, using its current state to plan an action and receiving an observation from the web environment.Its trajectory records planning thoughts, seeking actions, and resulting observations; the agent ultimately gathers unique entities.Source: §2.2 · pp. 3–4

2.3 Quantifying Information Collection and Efficiency

  • The performance framework measures both information collected and efficiency per action, with ISR tracking retrieved target entities and ISE reflecting collection efficiency.The paper states that ISE becomes a more stable and reliable metric as the number of target entities grows.Source: §2.3 · p. 4

3 Method

WebLeaper uses tree-structured tasks to include more target entities, then constructs trajectories and rewards that favor accurate, efficient information seeking. Its task variants expand from single-source trees to multi-source unions and reversed reasoning flows.

3 Method

  • Tree-structured tasks accommodate more target entities, addressing the limited training coverage and ISE measurement bias associated with small target sets.WebLeaper pairs task synthesis with trajectory construction.Source: §3 · pp. 4–5

3.1.1 Version-I: Basic

  • Basic tasks represent entities as a tree: table-derived roots connect to second-layer entities and their attribute-bearing third-layer entities.Each second-layer entity and its associated attributes form a subtree, and the subtrees make up the reasoning tree.Source: §3.1.1 · pp. 5–6

3.1.2 Version-II: Union

  • Union combines trees whose subtrees share relations, identifying maximal unions to avoid enumerating every possible combination.An LLM uses common features in the selected subtrees to synthesize questions requiring intermediate entity sets and their intersection.Source: §3.1.2 · p. 6

3.1.3 Version-III: Reverse-Union

  • Reverse-Union fuzzes an anchor entity into descriptive clues, then uses a pivot attribute to direct a search across unified trees for target entities.This reversed workflow is designed to prevent solving the task through simple keyword searches.Source: §3.1.3 · pp. 6–7

3.2 Information-Guided Trajectory Construction

  • Trajectory filtering retains examples only when ISR exceeds a coverage threshold and ISE exceeds an efficiency threshold.ISE counts target entities found in Visit actions, excluding Search results because those entities are less precise and may be updated by Visit.Source: §3.2 · p. 7

3.3 Reinforcement Learning with Hybrid Reward Systems

  • For entity-intensive tasks, WebLeaper uses a granular semantic F-score reward, while retaining existing rewards for public benchmark tasks and optimizing the policy with GRPO.Semantic scoring accommodates minor entity-name variations; the hybrid signal provides fine-grained feedback while remaining compatible with established evaluation protocols.Source: §3.3 · pp. 8–9

4 Experiments

The experiments test WebLeaper across five information-seeking benchmarks, comparing task-synthesis variants, trajectory filters, and training settings. The reported results show gains in benchmark performance and search efficiency, including after reinforcement learning.

4.1 Setup

  • WebLeaper leads mainstream open-source agents across five benchmarks; GAIA uses a 103-sample text-only validation subset, while other benchmarks use their complete test sets.The base configuration combines WebLeaper data with 5,000 WebSailor-V2 examples; the comprehensive setting mixes WebLeaper data into a larger corpus for SFT and RL.Source: §4.1 · p. 10, §4.2 · p. 11

4.2 Overall Performance

  • 73.2 on GAIA, 38.8 on BrowseComp, and 72.0 on xbench-DeepResearch; WebLeaper leads open-source agents across five benchmarks and tops WideSearch Success Rate and Item F1 in the comprehensive setting.In the base setting, it is comparable to or surpasses Claude-4-Sonnet and OpenAI-o3 on benchmarks other than BrowseComp and WideSearch, and outperforms Kimi-K2-Instruct-1T on BrowseComp.Source: §4.2 · p. 11

4.3 Capability Gains Induced by Entity-Intensive Task Synthesis

  • Union improves average performance by +3.26 over WebSailor-V2-5k, while Basic drops across three datasets and Reverse-Union brings widespread gains over Union.Basic tasks can invite shortcut learning; Union combines heterogeneous sources, while Reverse-Union adds reasoning complexity that supports planning and decision-making.Source: §4.3 · pp. 11–12

4.4 Impact of Information-Guided Trajectory Construction

  • ISR+ISE performs best on GAIA and BrowseComp, while the three trajectory-filtering strategies are comparable on WideSearch within the margin of variance.The authors suggest precision and efficiency constraints complement each other on more complex browsing tasks; on WideSearch, the choice of filter appears less consequential.Source: §4.4 · p. 12

4.5 Joint Gains in Efficiency and Effectiveness

  • WebLeaper improves effectiveness while using fewer average action rounds on WideSearch and BrowseComp, and reduces operational cost while improving effectiveness on GAIA and xbench-DS.The authors characterize this as a joint improvement in information-seeking efficiency and task performance compared with the baseline.Source: §4.5 · p. 12, Figure 5 · p. 12

4.6 Reinforcement Learning using WebLeaper

  • After reinforcement-learning fine-tuning, the model surpasses the SFT-only baseline across all benchmarks.The reward curve rises steadily through the evaluated 135 steps; the experiment stopped when web-access resources were exhausted.Source: §4.6 · p. 13, Table 3 · p. 13, Figure 6 · p. 13

5 Related Work

The related work places WebLeaper among information-seeking agents trained through fine-tuning and distinguishes its focus on entity-rich data from prior approaches emphasizing task difficulty or planning.

5.1 Information Seeking Agent

  • Information-seeking agents are grouped into model fine-tuning, agent-architecture advances, and multi-agent systems; this work aligns with the fine-tuning stream.Source: §5.1 · pp. 13–14
  • Prior training often targets correctness on single-fact answers, whereas this work treats the number of answer entities as a dimension of completeness and efficiency.Source: §5.1 · p. 14

5.2 Agent Data Synthesis

  • Prior information-seeking data synthesis emphasizes multi-step reasoning or long-horizon planning; this work instead centers on QA data with high entity-level complexity.The authors present semantic richness as complementary to those task-difficulty approaches.Source: §5.2 · p. 14

6 Conclusion

The conclusion presents WebLeaper as a way to address inefficient information seeking through entity-intensive tasks and accurate, efficient trajectories. It reports consistent gains across five benchmarks.

6 Conclusion

  • WebLeaper formulates information seeking as tree-structured reasoning and uses Basic, Union, and Reverse-Union variants to construct entity-intensive tasks.Its trajectory curation uses ISR and ISE to retain solutions that are accurate and efficient.Source: §6 · p. 14
  • WebLeaper consistently improves performance across five challenging benchmarks, supporting search efficiency as a lever for improving information-seeking agent capabilities.Source: §6 · p. 14

A.1 Declaration on the Use of LLMs

  • The authors report using LLMs for language-related manuscript assistance and academic research, with substantive content independently authored and reviewed.Source: §A.1 · p. 19

A.2 Proof of Proposition 1

The proof establishes that ISE variance decreases in proportion to the inverse of the number of required entities, under an i.i.d. assumption for entity-discovery steps. It uses the Delta method to approximate the variance.

A.2 Proof of Proposition 1

  • Var(ISE) = O(1/n): the proof states that ISE becomes more stable as the number of required entities grows.Source: §A.2 · p. 19
  • The derivation assumes the steps needed to discover each required entity are independent and identically distributed, with common mean and variance.Source: §A.2 · p. 19
  • The proof defines X as the average steps to discover one required entity and applies the Delta method to the ISE function f(X) = 1/X.Source: §A.2 · p. 19

A.3 Data Statistics

The training-data statistics show that many samples contain at least 100 entities. The paper presents this complexity as important for efficiency measurement and overall performance.

  • A significant portion of training samples contains at least 100 entities.The paper describes this entity count as underscoring the dataset's inherent difficulty.Source: §A.3 · p. 19
  • The authors state that this entity complexity is crucial for robust efficiency measurement.Source: §A.3 · p. 19
  • Figure 7 presents the distribution of the training data.Source: §A.3 · p. 19, Figure 7 · p. 20

A.4 Data Cleaning and Basic Task Construction

The data pipeline cleans Wikipedia tables and converts their rows into hierarchical reasoning trees. The tree organizes query entities, identifiable second-layer entities, and their associated attributes.

  • A basic task represents entities in a compact hierarchy, with each subtree grouping information associated with a common query concept.Source: §A.4 · p. 20
  • The pipeline starts with approximately 2 million crawled Wikipedia tables and applies multi-stage cleaning because raw web tables are often noisy and inconsistent.Source: §A.4 · p. 20
  • Filtering removes tables outside the 10–200-row and 3–20-column ranges, semantically irrelevant columns, and tables with substantial formatting errors.Source: §A.4 · p. 20
  • The pipeline retains isomorphic tables with shared headers and structure, supporting identification of common subtree structures for Union.Source: §A.4 · p. 20
  • The resulting clean tables have meaningful fields and multiple rows, each of which can be transformed into a subtree.Source: §A.4 · p. 20
  • The tree uses title or caption entities as its root, a unique-valued key column for second-layer subtree roots, and remaining column values as leaves.Source: §A.4 · pp. 20–21

A.5 Maximal Union Algorithm for Task Synthesis

The Union algorithm finds maximal groups of compatible reasoning trees by recasting their shared relations as maximal bicliques in a bipartite graph. It then filters candidates by size and semantic compatibility.

A.5 Maximal Union Algorithm for Task Synthesis

  • The algorithm reframes the search for compatible reasoning trees as maximal biclique enumeration to discover maximal union groups.Source: §A.5 · p. 21

Problem Formulation

  • The bipartite graph links each base tree to its subtree relations; a maximal biclique pairs trees with their common relations.Source: Problem Formulation · p. 21
  • Maximality means no tree or relation can be added without breaking the biclique property, and this formulation is computationally efficient compared with exhaustive search.Source: Problem Formulation · p. 21

Algorithm and Implementation Details

A.6 Detailed Examples of Task Synthesis

The examples illustrate three task variants built from reasoning trees: Basic retrieves attributes from one table, Union integrates entities across trees, and Reverse-Union begins from clues that obscure an anchor entity.

A.6.1 Version-I: Basic

  • Basic builds a hierarchical task from one table, using its title and constraints for scope, a key column for entities, and remaining columns for their attributes.The example asks for Nobel literature laureates from 1980 to 1990 and their names, countries, award years, and genders.Source: §A.6.1 · p. 23

A.6.2 Version-II: Union

  • Union requires relational integration across distinct reasoning trees, identifying target entities in both prize-winner sets and retrieving requested attributes.The example intersects Nobel Literature laureates and Booker Prize winners by name.Source: §A.6.2 · p. 23

A.6.3 Version-III: Reverse-Union

  • Reverse-Union obscures the query's entry point: the agent deduces an anchor from descriptive clues, then uses one of its attributes to find related target entities.In the example, clues identify William Golding; his nationality becomes the pivot for finding authors who share it and won both prizes.Source: §A.6.3 · p. 24

B Tool Call Analysis

WebLeaper uses many tool calls, often several dozen per instance and sometimes more than 50, reflecting intensive interaction and exploration.

B Tool Call Analysis

Loading 2510.24697v1…