---
title: 'WebLeaper: Efficient Web-Agent Training'
url: https://www.emergentmind.com/topics/webleaper
type: topic
---

# WebLeaper: Efficient Web-Agent Training

Searching arXiv for "WebLeaper" and closely related web-agent papers to ground the article.
WebLeaper is a data-centric training framework and agent design for information-seeking web agents, introduced in “WebLeaper: Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking” [2510.01524]. It addresses a specific weakness of contemporary large-language-model web agents: they often improve retrieval depth without improving search efficiency, because their training tasks are entity-sparse and provide weak signals about efficient information collection. WebLeaper reformulates information seeking as a tree-structured reasoning problem, synthesizes high-coverage tasks from curated Wikipedia tables through three variants—Basic, Union, and Reverse-Union—and trains on trajectories filtered for both correctness and efficiency. In the reported experiments, this data-centric design yields simultaneous gains in effectiveness and tool-call efficiency on BrowserComp, GAIA, xbench-DeepSearch, WideSearch, and Seal-0 [2510.01524].

## 1. Problem Setting and Data-Centric Premise

WebLeaper is positioned against the view that better web-agent performance follows primarily from deeper search or more powerful planners. Its starting observation is that current information-seeking agents often waste many actions on irrelevant pages, redundant query reformulations, and long detours, and that the fraction of valid actions in preliminary analysis peaks around approximately \(0.04\), corresponding to roughly one effective action per 25 tool calls [2510.01524]. The framework attributes this inefficiency to the sparsity of target entities in prevailing training tasks.

Three consequences of entity sparsity are central to the framework. First, sparse tasks provide poor supervision for efficiency learning, because most actions in a long trajectory gather no labeled intermediate information. Second, efficiency metrics become noisy when there are only a few targets, which weakens their usefulness as optimization signals. Third, sparse tasks permit brittle shortcuts, such as lucky keyword matches, rather than systematic multi-step entity collection. WebLeaper therefore shifts the emphasis from merely increasing reasoning depth to constructing “info-rich” tasks in which many intermediate and final entities are explicitly relevant to the solution [2510.01524].

A common misconception is that search inefficiency is mainly a runtime planning defect. WebLeaper instead treats it as a training-distribution defect: if the supervision signal does not densely reward information acquisition, the policy cannot reliably learn efficient search behavior. This premise explains why the framework concentrates on task construction and trajectory curation rather than on a novel inference-time architecture.

## 2. Formalization of Information Seeking

WebLeaper defines an information-seeking task as
$$
\mathcal{T} = \langle q, R \rangle,
$$
where \(q\) is a natural-language question and \(R \subset \mathcal{E}\) is the set of target entities required by the question. Importantly, \(R\) includes not only final answer entities but also intermediate entities required by the reasoning process. This broad definition of relevance is what makes the supervision signal substantially denser than in conventional answer-only formulations [2510.01524].

The agent follows a ReAct-style interaction pattern. A trajectory is written as
$$
\mathcal{H}_T = (q, \tau_1, \alpha_1, o_1, \tau_2, \alpha_2, o_2, \ldots, \tau_T, \alpha_T, o_T),
$$
where \(\tau_i\) is the thought text at step \(i\), \(\alpha_i\) is the action, and \(o_i\) is the observation. Across the trajectory, the agent accumulates an observed entity set \(O\).

WebLeaper uses two core metrics. The first is Information-Seeking Rate, defined as
$$
\mathrm{ISR} = \frac{|R \cap O|}{|R|},
$$
which measures recall over the target entity set. The second is Information-Seeking Efficiency,
$$
\mathrm{ISE} = \frac{|R|}{T},
$$
which measures entities acquired per step. ISR captures completeness; ISE captures cost-normalized productivity [2510.01524].

The framework also gives a variance argument for why entity-rich tasks are preferable. If \(X_i\) is the number of steps needed to discover the \(i\)-th new entity, with
$$
T = \sum_{i=1}^{n} X_i, \qquad \mathrm{ISE} = \frac{n}{T} = \frac{1}{\overline{X}},
$$
then, under the stated i.i.d. assumption and a delta-method approximation,
$$
\mathrm{Var}(\mathrm{ISE}) = \mathcal{O}\!\left(\frac{1}{n}\right).
$$
As the number of target entities \(n\) increases, the efficiency estimate becomes more stable. This gives a formal justification for WebLeaper’s emphasis on tasks with large target sets rather than sparse answer-only supervision [2510.01524].

## 3. Tree-Structured Task Synthesis

WebLeaper formulates information seeking as a tree-structured reasoning problem. Each task is represented by a tree whose root is a question entity, whose second layer consists of primary entities, and whose third layer contains attribute or related entities connected by typed relations. A second-layer node together with its attached third-layer entities forms a subtree \(S_{i,j}\), and \(\mathrm{Rel}(S_{i,j})\) denotes the set of relation types present in that subtree [2510.01524].

The source material for these trees is a large table corpus. WebLeaper crawls approximately 2 million Wikipedia tables and performs multi-stage cleaning, including filtering by size, removing trivial or structural columns, and keeping structurally homogeneous groups of isomorphic tables. This produces a table collection suitable for systematic synthesis of high-coverage tasks [2510.01524].

The framework defines three task variants.

| Variant | Construction | Role in training |
|---|---|---|
| Basic | Single-table reasoning tree | High coverage, structurally simple |
| Union | Multi-table composition via shared relations | Higher search complexity and cross-source integration |
| Reverse-Union | Union plus reverse reasoning from descriptive clues | Highest complexity and reduced shortcutting |

Basic tasks are built from a single cleaned table. The framework extracts a root entity from the table title or caption, asks an LLM to identify the primary-key column, uses its values as second-layer entities, and treats the other columns as third-layer attributes. A question is then synthesized from this structure. Because all relevant rows and attributes are targets, a Basic task is already entity-dense, but it remains relatively simple and often single-source [2510.01524].

Union tasks combine multiple reasoning trees. Let \(\mathcal{T}_{\text{base}} = \{T_1,\dots,T_N\}\), and for each tree define
$$
\mathrm{Rel}(T_i) = \bigcup_j \mathrm{Rel}(S_{i,j}).
$$
WebLeaper builds a bipartite graph between trees and relation types, and searches for maximal bicliques \((\mathcal{U}, \mathcal{V})\) satisfying
$$
\forall T_i \in \mathcal{U},\;\mathcal{V} \subseteq \mathrm{Rel}(T_i), \qquad |\mathcal{U}| \ge k_{\min},\ |\mathcal{V}| \ge m_{\min}.
$$
Each resulting union group supports cross-table questions that require collecting entity sets from multiple trees and combining them, typically through intersection-like reasoning [2510.01524].

Reverse-Union tasks further increase difficulty by reversing the initial reasoning direction. Instead of naming an anchor entity directly, the question begins with descriptive clues derived from its third-layer attributes. The agent must infer the anchor second-layer entity, extract one of its attributes as a pivot, and then perform a union-style search over related trees. This construction is designed to reduce keyword shortcutting and force a more structured search policy [2510.01524].

An important ablation confirms the hierarchy among these variants. When mixed with 5k WebSailor-V2 examples, Basic-5k reduces average benchmark performance by \(7.64\), whereas Union-5k improves average performance by \(3.26\), and Reverse-Union-10k improves it by \(4.34\). The paper therefore treats Basic as insufficient on its own and regards Union and especially Reverse-Union as the critical sources of transfer [2510.01524].

## 4. Trajectory Construction and Filtering

WebLeaper uses a tool-augmented agent with two tools: `Search(queries, filter_year)` and `Visit(urls, goal)`. `Search` returns top URLs and snippets for Google-like queries; `Visit` opens selected URLs and returns summarized paragraphs. Trajectories are generated with a Qwen3-30B-A3B-Thinking-2507 agent under ReAct prompting, using these tools to solve the synthesized tasks [2510.01524].

The framework then filters the generated trajectories. Coverage is assessed by ISR, and efficiency is assessed by ISE, but the implementation distinguishes the observation sources used for each. ISR counts target entities found through both `Search` and `Visit`, whereas ISE uses only entities obtained in `Visit` actions, on the rationale that search snippets are weaker evidence and are usually superseded by visited-page content [2510.01524].

The retention rule applies thresholds to both metrics:
$$
\mathrm{ISR} > \alpha, \qquad \mathrm{ISE} > \beta,
$$
with typical values \(\alpha = 0.3\) and \(\beta = 0.1\). Thus, a retained trajectory must both recover a nontrivial share of the target entity set and do so at a reasonable entity-per-step rate. This dual filtering removes trajectories that are complete but bloated as well as trajectories that are efficient but incomplete [2510.01524].

The framework evaluates ISR-only filtering, ISE-only filtering, and joint ISR+ISE filtering. On GAIA and BrowseComp, ISR+ISE yields the strongest downstream performance; on WideSearch, the three strategies are closer, but the joint criterion remains the default. This indicates that WebLeaper’s central training signal is not raw success alone but a coupled measure of coverage and search economy [2510.01524].

## 5. Training, Reward Design, and Reported Performance

WebLeaper is trained in two settings. In the base setting, WebLeaper task data are mixed with 5k WebSailor-V2 synthetic deep-search examples, and supervised fine-tuning is performed on the resulting trajectories. In the comprehensive setting, WebLeaper data are injected into the larger Tongyi-DeepResearch-30B-A3B training pipeline at both the SFT and RL stages [2510.01524].

For reinforcement learning, the framework defines a hybrid reward system. On legacy data it uses the original reward, while on WebLeaper tasks it uses a soft entity-level \(F\)-score. The soft recall and precision are
$$
R_c = \frac{1}{|R|} \sum_{e_r \in R} \max_{e_o \in O} s(e_o, e_r),
$$
$$
P = \frac{1}{|O|} \sum_{e_o \in O} \max_{e_r \in R} s(e_o, e_r),
$$
where \(s(e_o, e_r) \in [0,1]\) is a semantic scoring function between observed and target entities. The WebLeaper reward is then
$$
\mathcal{R}_{\text{WebLeaper}} = (1 + \omega^2)\,\frac{P \cdot R_c}{\omega^2 P + R_c},
$$
and the hybrid reward selects \(\mathcal{R}_{\text{WebLeaper}}\) for WebLeaper tasks and \(\mathcal{R}_{\text{legacy}}\) otherwise. Optimization uses a GRPO objective with group-standardized advantages and PPO-style clipping [2510.01524].

The reported benchmark results are strong in both the base and comprehensive settings. In the base setting, WebLeaper-Union reaches BrowseComp \(22.1\), GAIA \(69.9\), xbench-DeepSearch \(62.3\), Seal-0 \(35.1\), and WideSearch \(4.0\) SR, \(22.2\) Row F1, \(34.5\) Item F1. WebLeaper-Reverse-Union reaches BrowseComp \(23.0\), GAIA \(67.0\), xbench-DeepSearch \(66.0\), Seal-0 \(37.2\), and WideSearch \(4.0\) SR, \(25.8\) Row F1, \(40.8\) Item F1 [2510.01524].

In the comprehensive setting, WebLeaper-Reverse-Union combined with the Tongyi corpus reaches GAIA \(73.2\), BrowseComp \(38.8\), xbench-DeepSearch \(72.0\), Seal-0 \(48.6\), and WideSearch \(4.0\) SR, \(31.0\) Row F1, \(48.8\) Item F1 [2510.01524]. An RL ablation further reports that SFT-only training yields BrowseComp \(37.8\), GAIA \(69.9\), xbench-DeepSearch \(69.0\), and WideSearch \(1.5\) SR, \(23.0\) Row F1, \(45.4\) Item F1, while SFT+RL improves these to BrowseComp \(38.8\), GAIA \(73.2\), xbench-DeepSearch \(72.0\), and WideSearch \(4.0\) SR, \(31.0\) Row F1, \(48.5\) Item F1 [2510.01524].

A notable empirical claim is that WebLeaper improves both effectiveness and efficiency simultaneously. In the reported scatter plots against WebSailor-V2, WebLeaper attains higher benchmark scores while using fewer average tool calls on WideSearch, BrowseComp, GAIA, and xbench-DeepSearch. The framework therefore presents itself not as a trade-off mechanism between cost and accuracy, but as a way of reshaping the training distribution so that the learned policy moves to a better point on the Pareto frontier [2510.01524].

## 6. Relation to Subsequent Work and Scope

Subsequent papers place WebLeaper within the competitive landscape of search agents. “OpenSeeker-v2” reports WebLeaper-30B-SFT at BrowseComp \(27.7\) and xbench \(66.0\), and WebLeaper-30B-RL at BrowseComp \(38.8\) and xbench \(72.0\), while noting that its own 30B SFT-only agent reaches \(46.0\) on BrowseComp and \(78.0\) on xbench [2605.04036]. This comparison situates WebLeaper as a strong data-centric baseline and as a reference point for later academic search-agent systems.

Later benchmark work also uses “WebLeaper-style systems” as a label for deployment-oriented web agents that must navigate real sites, execute multi-step workflows, and complete end-to-end tasks rather than merely reach a page [2607.06118]. Although WebLeaper itself is centered on search with `Search` and `Visit` tools rather than full browser interaction, this later usage reflects its broader influence as a model of information-seeking web-agent training.

The framework’s scope is also delimited by its design choices. It relies on curated Wikipedia tables, entity-rich tasks, and LLM-as-a-judge components for both semantic matching and final evaluation [2510.01524]. This suggests that WebLeaper is most naturally aligned with tasks where relevant intermediate entities can be made explicit and abundant. A plausible implication is that its efficiency signal is strongest in settings where \(|R|\) is large and structurally organized, and weaker in tasks whose target set is intrinsically small or whose relevant state is not easily expressible as an entity set.

Within the literature on web agents, WebLeaper’s distinctive contribution is therefore not a new planner, browser substrate, or tool interface, but a training framework that treats efficient information acquisition as a first-class supervised and reinforced objective. Its lasting significance lies in the proposition that web-agent efficiency can be improved by redesigning the task and reward geometry of training, rather than by relying only on larger models or deeper search [2510.01524].

Source: https://www.emergentmind.com/topics/webleaper