Nested Browser-Use Learning for Agentic Information Seeking

1 Introduction

• Search snippets and single-URL fetches provide only partial web access, missing dynamic content, forms, multi-step navigation, and page-internal functionality.Such information is not reliably surfaced by search engines or recoverable through one static URL fetch.Source: §1 · p. 2
• Browser-use lacks a widely adopted action standard, while complex controls and redundant page outputs challenge ReAct agents, sometimes exceeding context limits.The paper frames a simple yet complete toolkit and efficient interaction paradigm as an open challenge.Source: §1 · p. 2
• NestBrowse’s minimal, functionally complete browser toolkit provides four core actions: search, visit, click, and fill.The toolkit is intended to enable comprehensive web information access with low tool complexity.Source: §1 · p. 2
• The Nested Browser-Use Framework separates outer-loop tool-integrated reasoning from inner-loop page exploration, controlling information flow into the agent context.A multi-task imitation-learning formulation jointly trains outer-loop reasoning and inner-loop evidence extraction in one IS agent model.Source: §1 · p. 2
• Across four highly challenging deep information-seeking benchmarks in both English and Chinese, NestBrowse-4B and NestBrowse-30B-A3B both exhibit consistently strong performance.Source: §1 · p. 2

2 Related Work

• Tool-integrated reasoning equips models with external tools and interleaves reasoning with tool invocation for tasks requiring information access or computation (Lin & Xu, 2025; Xue et al., 2025).Prior work identifies interface design as central and favors minimal, low-complexity toolkits to reduce decision burden and support efficient information flow (Shen, 2024; Liu et al., 2024b; Patil et al., 2024; Shen et al., 2023; Qin et al., 2024; Deng et al., 2023; Liu et al., 2023).Source: §2 · p. 2
• Deep IS tasks start from vague prompts and require agents to infer latent clues, entities, and relations through hypothesis generation, evidence gathering, and verification (Wu et al., 2025b; Wei et al., 2025; Yang et al., 2018; Li et al., 2025c; Li et al., 2025a).The passage contrasts this with conventional multi-hop QA, where entities and intermediate clues are provided.Source: §2 · pp. 2–3
• Prior web-based IS agents use ReAct-style tool invocation, with effective interaction abstractions and information-acquisition strategies remaining key challenges (Wu et al., 2025b;a; Qiao et al., 2025; Team, 2025c).Source: §2 · p. 3
• NestBrowse addresses these challenges with browser-based interactions intended to acquire more comprehensive and better-structured web information for complex deep IS tasks.Source: §2 · p. 3

3.1 Minimally Complete Browser Toolkit

• The four-tool toolkit combines search and visit for static information with click and fill for dynamic information, balancing access coverage against tool complexity.The authors describe these four tools as completing the information-access pathway while adding only two actions to standard search and visit.Source: §3.1 · p. 3
• The toolkit’s page content is text-only and does not incorporate visual information.This is an explicit scope boundary of the browser setup.Source: §3.1 · p. 3

3.2 Nested Browser-Use Framework

• The framework excludes scrolling and in-page search because they limit exposure per page read without improving goal-directed information acquisition.Its stated aim is to return only the subset of page information needed for the task goal.Source: §3.2 · p. 4
• Page transitions trigger an inner loop that explores page segments for the outer loop’s goal, returning accumulated goal-relevant content instead of raw page text.The outer loop continues reasoning with this compact response.Source: §3.2 · pp. 4–5

3.3 Multi-Task Imitation Learning

• SailorFog-QA-V2 is the primary training task because simple single- or multi-hop QA sets may not require substantive browsing or reasoning.The selected QA set is designed to elicit complex browsing and multi-step reasoning.Source: §3.3 · p. 5
• Rejection sampling retains trajectories that meet quality criteria, including valid formatting, executable tool calls, and correct final answers.Intermediate reasoning and browsing steps receive no additional rejection criteria, preserving supervision diversity and avoiding brittle manual constraints.Source: §3.3 · pp. 5–6
• The framework trains both outer-loop reasoning with tool use and inner-loop extraction of goal-relevant page evidence through a weighted multi-task objective.The outer objective supervises agent trajectories; the inner objective trains evidence extraction from page segments.Source: §3.3 · p. 6
• The method organizes learning around both final-answer generation and incremental maintenance of a temporary workspace containing relevant evidence.These are the outer- and inner-loop objectives induced by the nested framework.Source: §3.3 · p. 6

4.1 Setup

• Evaluation covers BrowseComp, GAIA, BrowseComp-zh, and XBench; performance is final-answer accuracy, judged with GPT-4.1, using GAIA’s 103-question text-only subset.The other benchmarks use their full datasets.Source: §4.1 · pp. 6–7

4.2 Overall Performance

• NestBrowse-30B-A3B delivers strong results across four deep information-seeking benchmarks, exceeding leading open-source agents and matching or outperforming several proprietary systems; NestBrowse-4B is competitive with many larger agents.Although trained solely on English data, NestBrowse generalizes well to the Chinese benchmarks.Source: Table 1 · p. 7, §4.2 · pp. 7–8

4.3 Analysis of Browser-Use Strategies

• Toolkit simplification and goal-relevant page extraction each improve performance, with extraction yielding larger gains; combining both produces the strongest results with a clear additive effect.The ablation compares strategies using the same GPT-OSS-120B agent model.Source: Table 2 · p. 8, §4.3 · p. 8

4.4 Context Efficiency of NestBrowse

• Figure 3 tracks average context length over tool-call turns for NestBrowse-30B-A3B, with gray bars showing the number of active trajectories remaining.Source: Figure 3 · p. 8

4.5 Analysis of Intra-Page Exploration

• NestBrowse improves both raw snapshot retention and goal-relevant extraction accuracy over the base agent, as measured by GPT-4.1.These metrics assess preservation of page snapshots for later interactions and alignment of extracted content with the task goal.Source: Figure 4 · p. 9, §4.5 · p. 9
• On the BrowseComp subset, weaker inner-loop models degrade outer-loop performance, while stronger ones improve it, supporting joint training of the two loops.Table 3 reports deep information-seeking performance for different inner-loop models on 100 examples.Source: §4.5 · p. 9, Table 3 · p. 9

4.6 Case Study: Beyond Static Web Access

• In a GAIA numerical-computation task, NestBrowse-30B-A3B uses an online calculator, reducing its reasoning burden and successfully completing the task.The case illustrates browser-use for accessing interactive web functionality beyond static page content.Source: Figure 5 · p. 9, §4.6 · p. 10

• NestBrowse combines a minimal yet complete browser toolkit with a nested browser-use paradigm for efficient deep-web information access.Source: §5 · p. 10
• The framework decouples agentic reasoning from intra-page exploration and trains these capabilities jointly.Source: §5 · p. 10
• NestBrowse achieves strong deep information-seeking performance across relatively small model scales.Source: §5 · p. 10

6 Limitations and Future Work

• NestBrowse intentionally restricts browser modeling to textual content.The stated aim is to isolate tool-abstraction and interaction-paradigm effects in a controlled, interpretable experiment.Source: §6 · p. 10
• Adding vision or audio would increase modeling and system complexity, making browser-use learning harder to disentangle from multimodal perception.Source: §6 · p. 10
• The authors propose multimodal information acquisition and reasoning as a future direction because non-textual modalities can convey important information during real-world browsing.Source: §6 · p. 10

A.1 Implementation Details: Prompts

• The appendix presents the prompts used in NestBrowse, starting with those for outer-loop agentic reasoning.Source: §A.1 · p. 15
• The outer-loop system prompt follows the standard ReAct-style function-calling format.Source: §A.1 · p. 15
• The outer-loop user prompt consists directly of the user query, with no additional instructional or guiding content.Source: §A.1 · p. 15

SYSTEM PROMPT (OUTER LOOP)

• The browser-use agent prompt calls for thorough, multi-source investigations using credible, diverse sources across broad and specialized queries.Source: SYSTEM PROMPT (OUTER LOOP) · p. 15
• The agent is instructed to present its definitive response inside <answer> tags and may call one or more functions to assist with the query.Source: SYSTEM PROMPT (OUTER LOOP) · p. 15
• The prompt provides function signatures within <tools> tags and requires each function call in a JSON object inside <tool_call> tags.Source: SYSTEM PROMPT (OUTER LOOP) · p. 15
• The inner-loop page-exploration prompt extracts goal-relevant webpage content into a predefined JSON format and incrementally maintains extracted information in a temporary workspace.Source: SYSTEM PROMPT (OUTER LOOP) · p. 15

SYSTEM PROMPT (INNER LOOP)

• The inner-loop system prompt requires one JSON object with the fields rational, evidence, and summary, wrapped in <useful_info> tags.Source: SYSTEM PROMPT (INNER LOOP) · pp. 15–16
• The prompt asks the agent to locate goal-relevant sections or data, extract relevant information with original context, and organize it in a concise summary.Source: SYSTEM PROMPT (INNER LOOP) · pp. 15–16
• The inner-loop user prompt combines webpage content with the specified goal to guide extraction and formatting.Source: SYSTEM PROMPT (INNER LOOP) · p. 16

USER PROMPT (INNER LOOP)

• The inner loop locates content relevant to the user’s goal, extracts supporting evidence, and organizes the result into a concise summary.The prompt asks to preserve relevant original context as fully as possible.Source: USER PROMPT (INNER LOOP) · p. 16
• The requested output format is JSON with rational, evidence, and summary fields.Source: USER PROMPT (INNER LOOP) · p. 16
• For multi-segment webpages, each iteration appends the previous workspace’s evidence and summary to the prompt, supporting incremental maintenance and reducing redundant extraction.The completed workspace returns to the outer loop for subsequent reasoning.Source: USER PROMPT (INNER LOOP) · p. 16

USER PROMPT INCREMENTAL (INNER LOOP)

• The incremental prompt scans webpage content for the user’s goal while considering existing evidence and summary.Source: USER PROMPT INCREMENTAL (INNER LOOP) · pp. 16–17
• Incremental processing must integrate existing and newly extracted information into one consolidated result without labeling which parts are new.Source: USER PROMPT INCREMENTAL (INNER LOOP) · p. 17
• The incremental variant requests JSON output with rational, evidence, and summary fields.Source: USER PROMPT INCREMENTAL (INNER LOOP) · p. 17