WebAggregator-32B: Scalable Web Aggregator

• WebAggregator-32B is a 32-billion parameter deep research agent built on the Qwen3-32B architecture and designed for automated, verifiable web data aggregation.
• It employs the Explore-to-Evolve paradigm to construct an evidence-rich, multi-domain dataset and synthesize complex aggregation logic from diverse online sources.
• Benchmark tests show its robust multi-hop reasoning and aggregation efficiency, narrowing the performance gap with advanced commercial agents.

WebAggregator-32B is a 32B-parameter foundation model and deep research agent designed for scalable, verifiable information aggregation from the live web. As the largest variant in the WebAggregator family, it is developed on top of the Qwen3-32B architecture and fine-tuned via agentic, supervised trajectories on a novel dataset constructed with the Explore-to-Evolve paradigm. WebAggregator-32B demonstrates leading aggregation capabilities on the GAIA-text and WebAggregatorQA benchmarks and establishes new approaches to automated agent training for multi-hop, aggregation-intensive research queries (Wang et al., 16 Oct 2025).

1. Model Architecture and Core Agent Framework

WebAggregator-32B employs the Qwen3-32B architecture, a decoder-only Transformer with the following specifications:

• Parameter count: ~32 billion
• Layers: 64 Transformer layers
• Hidden size: 8,192; feed-forward inner dimension: 32,768
• Attention: 32 attention heads per layer
• Context window: up to 128,000 tokens

The agent utilizes the SmolAgents framework with a ReAct-style scaffold. This approach maintains the core Transformer unchanged while augmenting it with structured, prompt-based tool-calling abilities for interacting with complex, dynamic environments—such as “Search,” “Visit,” “Click,” “FileRead,” and “ImageCaption.” No bespoke aggregation modules are added; aggregation arises from conditioning the model on the observed trajectory of tool calls, extracted evidence, and aggregation hints.

2. Explore-to-Evolve Training Paradigm

The Explore-to-Evolve methodology underpins both data construction and agent training:

• Proactive Web Exploration: Starting from 5,000 anchor queries spanning 11 domains (e.g., Finance, Sport, Medical, Politics), the agent explores up to 160,000 URLs with a strict 30-step per-task budget and a minimum of 7 page visits per task. Tool use is adaptive: an average of 3–5 tools per trajectory.
• Corpus Creation: This process yields an evidence-rich, web-grounded corpus of approximately 54,000 unique URLs, including dynamic web pages, PDF/CSV files, and images.
• Automatic Aggregation Logic Synthesis: Given high-level aggregation guidance (12 operation subtypes; see Section 3), the agent composes, refines, and executes chains of operations (e.g., “Statistic→standard deviation,” “Correlate→Pearson coefficient,” “Trend Forecasting→exponential smoothing”) directly on the gathered evidence.

This paradigm enables collection of high-quality agent trajectories for robust supervised fine-tuning.

3. WebAggregatorQA Dataset and Aggregation Logic

The WebAggregatorQA dataset—central to model training—contains 9,883 training tasks and 200 human-verified test tasks, sourced from over 54,000 URLs across 11+ domains. Notable properties include:

• Trajectory Length: Average synthesis involves approximately 15 agent steps per task.
• Aggregation Requirements: Tasks require at least three compositionally distinct aggregation operations, ensuring complexity.
• Diversity and Balance: Aggregation types are selected for domain coverage and diversity—rare types are upsampled via annotation, and domain balance is curated by GPT-4.1.
• Operation Categories and Subtypes:

| Operation Category | Subtypes | Illustrative Operation Example | |---------------------|----------------------------------------------|------------------------------------------------------| | Element | Retrieve, Inverse, Math | “Which actor won X years after…?” | | Set | Filter, Existence, Compose | “How many of the top 100 cities have decreased?” | | Temporal Reasoning | Change, TempCalc | “Who has the longest nomination-span?” | | Scientific Analysis | CompIntensive, Predict, Statistic, Correlate | “Compute std. dev. of playoff win %.” |

• Formal aggregation logic: Typical operation for standard deviation $$\sigma = \sqrt{\frac{1}{n}\sum_{i=1}^n (x_i - \bar x)^2}$$ for Pearson correlation

$$r = \frac{\sum_{i=1}^n(x_i - \bar x)(y_i - \bar y)}{\sqrt{\sum_{i=1}^n(x_i-\bar x)^2}\,\sqrt{\sum_{i=1}^n(y_i-\bar y)^2}}$$

4. Supervised Fine-Tuning and Evaluation Protocol

Each agent trajectory is recorded in the format $(Q, a_1, o_1, \dots, a_n, o_n, A)$, where $Q$ is the query, $a_i$ tool actions, $o_i$ corresponding observations, and $A$ is the final answer. Critical setup details:

• Masked training: During fine-tuning, all future actions/observations are masked. The model is conditioned only on previous steps and learns to predict the next action and the final answer.
• Trajectory Quality: Training is restricted to 6,184 high-quality trajectories filtered via rejection sampling, output validation, and the retention of exceptions.
• Benchmarks: Evaluation is performed on GAIA-text (103 cases) and WebAggregatorQA (159 human-verified test cases), with automatic GPT-4.1-based answer validation.
• Metrics: Primary measures are pass@1 (exact match with the reference answer) and pass@3.

| Model | GAIA-text Avg (pass@1, %) | WebAggQA Avg (pass@1, %) | |------------------------------- |--------------------------:|-------------------------:| | GPT-4.1 (zero-shot agent) | 43.7 | 25.8 | | Claude-3.7-sonnet (agent) | 60.2 | 28.3 | | WebAggregator-32B (Qwen3-32B) | 56.3 | 26.4 |

On GAIA-text, WebAggregator-32B exceeds GPT-4.1 performance by 12.6 percentage points; on WebAggregatorQA it approaches the level of advanced commercial agents. Pass@3 on GAIA-text is 69.9%, >10 points above GPT-4.1.

Across other benchmarks (WebWalkerQA, XBench), even smaller variants (7B/8B) of WebAggregator outperform predecessor models such as WebDancer and WebSailor.

5. Case Study: Multi-Step Real-World Aggregation

The multi-hop nature and explicit aggregation in WebAggregator-32B is demonstrated by queries such as:

Example: “Among all Cleveland Cavaliers head coaches who have won at least one playoff game with the team, what is the standard deviation of their playoff win percentages?”

• Exploration Steps:

1. Search for head coaches.
2. Visit ESPN and extract list.
3. For each coach, visit playoff record pages and extract wins/losses.

• Aggregation/Evolution Steps:

1. Compute win % for each coach: $$\, \text{win\%} = \frac{\text{Wins}}{\text{Wins+Losses}} \,$$
2. Compute standard deviation:

   $\sigma = \sqrt{\tfrac1n\sum (p_i-\bar p)^2}$

• Answer: “0.127”
• References: Multiple URLs traversed via agentic tool use.

This illustrates non-trivial synthesis over heterogeneous evidence, multi-step reasoning, and numeric/statistical computation from the raw web.

6. Contributions, Limitations, and Future Directions

Key Contributions:

• First scalable, automated data pipeline for constructing training and evaluation tasks that emphasize both web exploration and verifiable aggregation.
• WebAggregator-32B narrows the performance gap to strong commercial agents on aggregation-centric web QA.
• Construction and validation of WebAggregatorQA, a robust benchmark that explicitly targets the aggregation bottleneck observed in state-of-the-art agents.

Limitations:

• Architectural choices forego dedicated aggregation modules; performance depends on generic Transformer capacity guided by tool-based prompting.
• The WebAggregatorQA test set is relatively small (159 annotated samples), limiting comprehensive coverage of real-world aggregation phenomena.
• Aggregation errors persist even when correct references are retrieved—a failure mode indicating deficiencies in stepwise reasoning and self-verification.
• Tasks dependent on real-time or rapidly shifting data challenge system stability; streaming and temporality-aware aggregation remain unsolved.

Future Directions:

This suggests that research may pivot to integrating specialized reasoning modules or hybrid retrieval–aggregation architectures. Expanding dataset scale, multimodal integration, and more granular self-critique/verification are likely areas of focus.

7. Significance in the Context of Deep Research Agents

WebAggregator-32B systematically advances the state-of-the-art for agent-based, open-domain research requiring aggregation, compositional reasoning, and synthesis across diverse document types. The Explore-to-Evolve paradigm, tool-augmented agent scaffolding, and dataset/benchmark contributions establish a reference methodology for measuring and improving deep research competencies in LLMs operating over the web. Nevertheless, closing the gap to human-level synthesis of heterogeneous and multi-step evidence remains a frontier challenge for future agent foundation models (Wang et al., 16 Oct 2025).