💻 computer science

BaRA: BFS-and-Reflection Web Data Collection Agent

This paper introduces BaRA, a web data collection agent that combines bounded breadth-first search with history-based self-reflection to outperform existing LLM-based methods in discovering links and extracting downloadable multimodal content from complex websites under a fixed interaction budget.

Original authors: Soojeong Lee, Joseph Lee, Yongseong Cho, Sunjae Kim, Youngwoo Moon, Kyungwoo Song

Published 2026-07-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Soojeong Lee, Joseph Lee, Yongseong Cho, Sunjae Kim, Youngwoo Moon, Kyungwoo Song

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to gather every single photo, video, and news article from a specific website to build a library. You have a team of intelligent robots (powered by Large Language Models) to do the job. However, these robots have a major problem: they are easily distracted. If they find a promising link, they might dive deep down a rabbit hole, missing huge sections of the site. Or, they might get stuck on a page that keeps loading new content, fail to grab the actual files, and just guess where the files might be.

The paper introduces BaRA (BFS-and-Reflection Agent), a new system designed to fix these mistakes. Think of BaRA not just as a robot, but as a strict project manager with a very specific rulebook.

Here is how BaRA works, broken down into simple concepts:

1. The "Breadth-First" Map (The Strategy)

Most web agents act like a tourist who sees a cool sign and immediately runs down that path, forgetting to look at the other streets. They get lost in the "deep" parts of a site and miss the main attractions.

BaRA uses a strategy called Bounded Breadth-First Search (BFS).

The Analogy: Imagine you are exploring a new city, but you have a strict rule: "Visit every street on the first block, then every street on the second block, before you go deeper."
The Benefit: Instead of diving deep into one alleyway and missing the whole neighborhood, BaRA systematically spreads out like a ripple in a pond. It ensures it visits a wide variety of pages before going deeper, guaranteeing it doesn't miss large chunks of the website.

2. The "Self-Correction" Loop (The Reflection)

Sometimes, even with a good map, things go wrong. A website might pop up a confusing ad, a video might fail to load, or the page might freeze. A normal robot might just give up and say, "I can't do this," or it might hallucinate (make up) fake links to pretend it succeeded.

BaRA has a Self-Reflection module.

The Analogy: Imagine a detective who tries to solve a puzzle. If they try a clue and it leads to a dead end, they don't just quit. They look at their notes, realize, "Oh, I tried to open the door, but it was locked. Next time, I'll try the window instead."
The Benefit: When BaRA fails to get a file or a link, it analyzes why it failed. It then rewrites its own instructions to try a different approach (like scrolling more or looking in a different spot) before giving up. It tries again, up to a set limit, turning failures into learning moments.

3. The "Quality Control" Check (The Validation)

Collecting a list of links is easy; collecting working links is hard. Robots often return links that look real but lead to broken pages or files that can't be downloaded.

BaRA adds a final Validation Step.

The Analogy: Before a delivery truck leaves the warehouse, a manager checks every box. If a box is empty or the address is fake, it's thrown out. BaRA doesn't just collect the URL; it actually tries to "download" the image or video to prove it works.
The Benefit: The final output only contains files that are guaranteed to be real and downloadable, not just guesses.

The Results: How Did It Do?

The researchers tested BaRA on two types of websites:

Synthetic Sites: 50 fake websites built specifically to test the system (like a training gym).
Real Sites: Three messy, real-world websites (Berkeley.edu, Imgur.com, and the Library of Congress) that are full of pop-ups, auto-loading videos, and confusing layouts.

The Findings:

Better Coverage: BaRA found more pages than the other robots because it didn't get distracted by deep, dead-end paths.
Better Recovery: When the other robots gave up on a tricky page, BaRA used its "reflection" to try again and succeed.
Real-World Wins: On the messy real-world sites, BaRA was the only system that could successfully recover video and image files from dynamic pages (like the endless scroll of Imgur) where other systems failed completely.

The Bottom Line

BaRA is a framework that makes web data collection more reliable by combining a systematic, wide-ranging search strategy with a smart, self-correcting mindset. It ensures that when a robot is sent to collect data, it doesn't just wander aimlessly or give up easily; it follows a plan, learns from its mistakes, and only brings back files that are actually real.

Problem Statement
Large language model (LLM)-based web agents have reduced the need for manual scripting in web data collection. However, when deployed on live websites, these agents frequently exhibit fragility: they often miss relevant pages, return incomplete multimodal outputs, or generate media URLs that are not directly downloadable. These failures are often "silent," meaning agents terminate without explicit errors while still failing to cover the necessary site scope or retrieve valid artifacts. Furthermore, existing benchmarks often conflate navigation success with extraction quality, making it difficult to isolate whether an agent failed to find a page or failed to extract data from a found page. The paper addresses the specific, under-emphasized setting of site-level data collection under a fixed interaction budget, where the goal is to cover a bounded set of internal pages and return verifiable, downloadable multimodal artifacts with provenance.

Methodology: The BaRA Agent
The authors propose the BFS-and-Reflection Agent (BaRA), a control layer designed to operate on top of existing browser-executed runtimes (specifically built on Browser-use). BaRA is engineered to address two common failure modes: premature commitment to a single deep branch (leading to poor coverage) and brittle failure after incomplete runs. The framework consists of three core components:

Bounded Breadth-First Search (BFS) Traversal:
Instead of allowing the agent to drift into deep, unstructured paths, BaRA employs a bounded BFS controller. Starting from a seed URL, the agent maintains a FIFO queue of URLs and depths. It systematically expands the frontier by:
- Normalizing candidate URLs and removing duplicates.
- Filtering out off-site links and non-content endpoints (e.g., login pages).
- Retaining at most $W$ unique child links per page in DOM order.
- Enforcing a maximum depth $D$ and a maximum number of visited pages $P$ .
  This ensures predictable coverage and generates an interpretable visitation record.
History-Based Self-Reflection:
Recognizing that web environments are brittle (due to dynamic DOMs, pop-ups, or timeouts), BaRA implements a reflection loop. If an attempt fails or yields incomplete outputs (e.g., empty text or missing media URLs), the system does not terminate. Instead, a reflection module analyzes the execution history and partial outputs to diagnose the failure type and point. It then rewrites the instruction for the next attempt (up to a retry limit $A$ ) to better align with the encountered website behavior (e.g., instructing the agent to continue scrolling or re-extract media).
Constrained Per-Page Extraction and Validation:
Once pages are discovered, BaRA applies a strict extraction protocol. The agent must stay on the target URL, perform bounded scrolling to reveal lazy-loaded content, and output raw text along with image and video URLs in a strict format. Crucially, the pipeline includes a download validation step where extracted media URLs are verified for downloadability before being returned as final artifacts.

Key Contributions
The paper outlines three primary contributions:

A Robustness-Oriented Framework: A collection framework that integrates bounded BFS traversal with history-based self-reflection under fixed interaction budgets, specifically designed to improve site-level coverage and artifact validity.
A Constrained Extraction Protocol: A per-page extraction and validation protocol that ensures returned multimodal artifacts are machine-processable and download-valid.
A Novel Evaluation Setup: An evaluation methodology that explicitly separates link discovery (Step 1) from download-valid extraction (Steps 2–3). This allows for a granular analysis of whether failures stem from navigation errors or extraction/validation errors across both synthetic and live web settings.

Experimental Results
The authors evaluated BaRA on 50 synthetic websites (with ground-truth reference sets) and three public real-world websites (berkeley.edu, imgur.com, loc.gov) featuring cluttered or dynamic layouts. The baselines included Pure LLM (non-interactive), SeeAct-Vision (vision-enabled), and Browser-use (standard browsing agent).

Link Discovery: BaRA achieved perfect accuracy (1.0000) across all benchmarks, reliably enumerating the target set of internal page URLs under the given budgets. In contrast, Browser-use dropped significantly on real websites (e.g., 0.6250 on berkeley.edu), and Pure LLM performed poorly across the board, often hallucinating non-existent hierarchies.
Multimodal Extraction: BaRA consistently outperformed baselines in text, image, and video extraction. The most significant gains were observed in download-valid image and video recovery. For instance, on the synthetic benchmark, BaRA improved image accuracy by +0.2200 and video accuracy by +0.2600 over Browser-use. On imgur.com, a highly dynamic site where all baselines failed to retrieve valid media, BaRA was the only method to achieve non-zero scores for image and video extraction.
Qualitative Analysis: Case studies revealed that while other agents might reach the correct pages, they often returned hallucinated URLs or missed lazy-loaded content. BaRA's combination of structured traversal and recovery logic allowed it to faithfully recover actual internal page URLs and the specific media sources used by the site generators.

Significance and Claims
The paper claims that BaRA demonstrates that coverage-oriented control (via bounded BFS) is critical for robust link discovery on real-world websites, often more so than stronger grounding alone. Furthermore, it argues that history-based self-reflection and download validation are necessary to convert site coverage into reliable, usable outputs. The authors emphasize that reaching the right pages is only the "first half" of the problem; the system must also preserve page context, expose lazy-loaded content, and return resources that are actually downloadable.

The work positions BaRA as a research prototype that highlights the need for explicit coverage control and artifact validation in web data collection. The authors modestly note that the framework still depends on LLM behavior and browser stability, and that future work should address richer interactions (tabs, modals) and streaming media. They also caution against misuse, recommending the tool be restricted to public pages and used in compliance with site policies.

1. The "Breadth-First" Map (The Strategy)

2. The "Self-Correction" Loop (The Reflection)

3. The "Quality Control" Check (The Validation)

The Results: How Did It Do?

The Bottom Line

More like this