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

TL;DR

WebLeaper tackles low search efficiency in LLM-based information-seeking agents by creating entity-rich training tasks and efficient solution trajectories. It formalizes information seeking as tree-structured reasoning and introduces Basic, Union, and Reverse-Union variants to densely populate target entities, paired with information-guided trajectory filtering via ISR and ISE. A Hybrid Reward System, combining granular F-score signals with legacy rewards and optimized through GRPO, yields improvements across five public benchmarks, showing strong joint gains in accuracy and efficiency. The work demonstrates that high-entity-density data and carefully designed training signals can substantially boost autonomous information gathering in web-enabled agents, with practical implications for scalable, efficient IS systems.

Abstract

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.

Figures (9)

• Figure 1: Results on comprehensive training setting. All WebLeaper scores are averaged over three runs. The metric of the first three figures are accuracy. "SR" denotes Success Rate on WideSearch.
• Figure 2: The distribution of valid actions of the agent based on the GPT model on our synthesized IS task. The valid actions are those seeking the correct target entities required by the question.
• Figure 3: An overview of WebLeaper. The reasoning structure is modeled as a tree. A root entity (question entity) connects to a set of second-layer entities. (a) Version-I (Basic) constructs a simple reasoning tree from a single information source. (b) Version-II (Union) creates a complex task by finding a maximal union between two trees that share a common set of relations within their subtrees (e.g., both have "has_nationality"). (c) Version-III (Reverse-Union) reverses the reasoning process. It provides fuzzed clues (third-layer entities) as question entities, forcing the agent to first deduce a second-layer anchor entity (an entity from the second layer), then other relevant subtrees.
• Figure 4: Ablation study results on information-guided trajectory construction strategies.
• Figure 5: Effectiveness and efficiency comparison between WebLeaper and WebSailor-V2.

Theorems & Definitions (2)

• Proposition 1: Variance of ISE
• proof