---
title: WebLeaper-30B-A3B
canonical_url: "https://www.modelscope.cn/models/iic/WebLeaper-30B-A3B"
md_url: "https://www.modelscope.cn/models/iic/WebLeaper-30B-A3B.md"
repository: iic/WebLeaper-30B-A3B
last_updated: 2025-12-24
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - qwen3_moe
architectures:
  - Qwen3MoeForCausalLM
library_name:
  - safetensors
  - pytorch
frameworks:
  - Pytorch
downloads: 64
stars: 2
---

# WebLeaper-30B-A3B

> WebLeaper-30B-A3B - iic 在 ModelScope 开源的模型。WebLeaper: Empowering Efficient, Info-Rich Seeking for Web Agents

- **Repository**: iic/WebLeaper-30B-A3B
- **Tasks**: text-generation
- **Downloads**: 64
- **Stars**: 2
- **Last updated**: 2025-12-24

Source: https://www.modelscope.cn/models/iic/WebLeaper-30B-A3B

---

# WebLeaper: Empowering Efficient, Info-Rich Seeking for Web Agents


💡 Introduction

  * **WebLeaper** is a data+training framework that makes web agents both effective and efficient at information seeking (IS).
  * **Key Idea:** Today’s IS agents waste many steps because training tasks contain sparse targets, so agents learn shallow, meandering search. WebLeaper fixes this with **entity-intensive tasks** and **efficiency-aware training**.
  * We cast IS as tree-structured reasoning and synthesize three data variants—**`Basic`**, **`Union`**, and **`Reverse-Union`**—to progressively increase entity density, cross-source reasoning, and anti-shortcut robustness.
  * We curate training data in two stages:
    1.  **SFT:** We filter trajectories with **Information-Seeking Rate (ISR)** for coverage and **Information-Seeking Efficiency (ISE)** for action economy, keeping only those that are accurate and fast.
    2.  **RL:** We introduce a **Hybrid Reward System** that provides a dense, granular signal for our entity-intensive tasks, optimizing the agent further with policy optimization.
  * Extensive experiments on **BrowseComp, GAIA, Seal-0, WideSearch, and xbench-DeepSearch** show consistent gains over strong open and proprietary baselines.



🚀 Highlights

  * **Entity-Intensive Tasks:** Pack many targets into compact contexts to teach efficient retrieval.
  * **Three Synthesis Modes:**
      * **`Basic`:** Single-table, high-density tasks.
      * **`Union`:** Multi-source, structure-aligned fusion for realistic synthesis.
      * **`Reverse-Union`:** “Fuzzed” anchors that force deduction before cross-source search—no keyword shortcuts.
  * **Efficiency-Aware SFT Curation:** Keep only SFT trajectories with high ISR and high ISE.
  * **Hybrid-Reward RL:** A novel, soft F-score-based reward ($\mathcal{R}_{WebLeaper}$) for entity-intensive tasks and **Group Relative Policy Optimization (GRPO)** to further boost performance.
  * **Strong Open-Source Results:** Strong improvements across 5 IS benchmarks with fewer actions per success.



-----

📦 Dataset & Task Design

WebLeaper constructs IS tasks from curated Wikipedia tables and cross-table unions:

  * **Tree-Structured IS:**
      * Root (question entity) $\rightarrow$ 2nd layer (key entities) $\rightarrow$ 3rd layer (attributes/linked entities)
      * Each 2nd-layer node + its attributes forms a subtree; tasks require retrieving final and intermediate entities.
  * **Variants**
    1.  **`Basic`:** Build a single-source tree from one well-formed table; dense targets in a constrained context.
    2.  **`Union`:** Detect maximal unions among trees that share relations (modeled as maximal biclique enumeration) to create multi-source synthesis questions.
    3.  **`Reverse-Union`:** Provide attribute-level clues to deduce a hidden anchor entity first, then pivot (e.g., nationality) to launch a union-style search.

> **Result:** Tasks that reward efficient exploration, resist keyword shortcuts, and stabilize metric estimation as the target count ($n$) grows.

-----

📐 Metrics: Measuring Coverage & Efficiency

  * **Information-Seeking Rate (ISR):** Fraction of required entities retrieved.
    $$
    \mathrm{ISR} = \frac{|R\cap O|}{|R|}
    $$
  * **Information-Seeking Efficiency (ISE):** Target entities discovered per action step.
    $$
    \mathrm{ISE} = \frac{n}{T}
    $$
  * **Stability:** As the number of targets $n$ increases, $\mathrm{Var}(\mathrm{ISE}) = \mathcal{O}(1/n)$, yielding reliable efficiency signals during training.

-----

🔍 Method Details

1.  **`Basic` (Single-Source, Dense)**

      * Mine large, homogeneous Wikipedia tables.
      * Root from table title; primary key columns $\rightarrow$ 2nd-layer; other columns $\rightarrow$ 3rd-layer attributes.
      * Build compact, high-coverage tasks that maximize valid actions.

2.  **`Union` (Multi-Source, Structured)**

      * Identify maximal unions between trees sharing relation sets (e.g., `has_nationality`, `has_name`).
      * Synthesize questions that require intersection/union across sources (e.g., “authors who won both Prize A and Prize B”).

3.  **`Reverse-Union` (Deduction $\rightarrow$ Expansion)**

      * Provide fuzzed clues at the attribute level to force anchor deduction (no direct keywords).
      * Use a pivot attribute (e.g., country) from the deduced anchor to launch a new `Union`-style search over other trees.

-----


## 🚩 Citation

If this work is helpful, please kindly cite as:

```bigquery
@misc{tao2025webleaper,
      title={Webleaper: Empowering efficiency and efficacy in webagent via enabling info-rich seeking}, 
      author={Tao, Zhengwei and Shen, Haiyang and Li, Baixuan and Yin, Wenbiao and Wu, Jialong and Li, Kuan and Zhang, Zhongwang and Yin, Huifeng and Ye, Rui and Zhang, Liwen and others},
      year={2025},
      eprint={2510.24697},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.24697}, 
}
```