nestbrowse-info-seeking

name	nestbrowse-info-seeking
title	Nested Browser-Use Learning for Agentic Information Seeking
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2512.23647
keywords	["information-seeking","web-agents","nested-architecture","efficiency"]
description	Scale information-seeking agents via nested architecture separating outer reasoning from inner page exploration. Minimal toolkit (search, visit, click, fill) handles dynamic web interactions, inner loop filters content before context injection, multi-task learning trains outer/inner jointly—enabling 4B models to match larger competitors.

Overview

Nested architecture decouples reasoning from page content extraction.

Core Technique

Nested Loop Structure:

# Outer loop: reasoning
reasoning = outer_agent.think(state)
tool_call = outer_agent.select_tool()

# Inner loop: extract relevant content
relevant_content = inner_agent.extract(page, goal)

# Return minimal content to reasoning

When to Use

Use when: Information-seeking agents, large-scale web interaction, efficiency critical.

References

• Nested outer/inner loop architecture
• Minimal browser toolkit
• Content filtering before context injection