Explore to Evolve: Scaling Evolved Aggregation Logic via Proactive Online Exploration for Deep Research Agents

TL;DR

This paper identifies a critical gap in web agents: robust information aggregation lags behind information seeking. It proposes an automated Explore to Evolve workflow that first conducts proactive online exploration and then automatically synthesizes aggregation logic to generate verifiable QA data, yielding the WebAggregatorQA dataset and the WebAggregator foundation models. Empirical results show WebAggregator-8B approaching GPT-4.1 on GAIA-text, while the 32B variant substantially surpasses GPT-4.1 on GAIA-text and nears Claude-3.7-sonnet on WebAggregatorQA, though the latter remains challenging, underscoring the difficulty of true aggregation. The work also introduces rigorous quality control, data leakage avoidance, and a multimodal, domain-diverse test set to drive progress toward web agents capable of deep analysis and synthesis across real-world web environments.

Abstract

Deep research web agents not only retrieve information from diverse sources such as web environments, files, and multimodal inputs, but more importantly, they need to rigorously analyze and aggregate knowledge for insightful research. However, existing open-source deep research agents predominantly focus on enhancing information-seeking capabilities of web agents to locate specific information, while overlooking the essential need for information aggregation, which would limit their ability to support in-depth research. We propose an Explore to Evolve paradigm to scalably construct verifiable training data for web agents. Begins with proactive online exploration, an agent sources grounded information by exploring the real web. Using the collected evidence, the agent then self-evolves an aggregation program by selecting, composing, and refining operations from 12 high-level logical types to synthesize a verifiable QA pair. This evolution from high-level guidance to concrete operations allowed us to scalably produce WebAggregatorQA, a dataset of 10K samples across 50K websites and 11 domains. Based on an open-source agent framework, SmolAgents, we collect supervised fine-tuning trajectories to develop a series of foundation models, WebAggregator. WebAggregator-8B matches the performance of GPT-4.1, while the 32B variant surpasses GPT-4.1 by more than 10% on GAIA-text and closely approaches Claude-3.7-sonnet. Moreover, given the limited availability of benchmarks that evaluate web agents' information aggregation abilities, we construct a human-annotated evaluation split of WebAggregatorQA as a challenging test set. On this benchmark, Claude-3.7-sonnet only achieves 28%, and GPT-4.1 scores 25.8%. Even when agents manage to retrieve all references, they still struggle on WebAggregatorQA, highlighting the need to strengthen the information aggregation capabilities of web agent foundations.

Figures (9)

• Figure 1: The Pass@1 performance of our WebAggregator models, tuned on the automatically constructed training resource, WebAggregatorQA, is comparable to or even exceeds that of GPT-4.1 on both GAIA-text and the more challenging WebAggregatorQA test set.
• Figure 2: The Explore to Evolve data construction pipeline of WebAggregatorQA. (1) Proactive Online Web Exploring gathers comprehensive information by interacting with the web environment through tools (more details in Figure \ref{['fig:web-explore']}). (2) Task Construction via Automatic Aggregation Logic Synthesis constructs QA pairs grounded on the explored knowledge by instantiating and evolving the high-level aggregation guidance into concrete operations, e.g., Statistic Analysis$\rightarrow$standard deviation. (3) Quality Control ensures the data quality and diversity.
• Figure 3: The distribution of domains in WebAggregatorQA, tasks categorized by the number of different tools involved during construction, and steps (an action-observation round) in data synthesis.
• Figure 4: Word cloud of aggregation operations extracted from the constructed tasks. In the Automatic Aggregation Logic Synthesis stage, the agent converts high-level guidance into concrete low-level operations to combine knowledge snippets into new conclusions. The illustrated task requires seeking knowledge by Search, Visit, Click, FileRead, and aggregations to derive the final answer.
• Figure 5: Samples from TaskCraft shi_taskcraft_2025, WebDancer wu_webdancer_2025, and WebShaper tao2025webshaper primarily evaluate basic information-seeking skills, such as Element -> Retrieve and Set -> Sets Composition for entity filtering. In contrast, the selected WebAggregatorQA samples demand significantly more complex information aggregation to derive final answers. Crucially, these diverse aggregation strategies are automatically evolved by agents, guided by high-level logics and accumulated knowledge during data construction, resulting in rich variability that reflects task-specific intricacies.