BaRA: BFS-and-Reflection Web Data Collection Agent

arXiv cs.AI Papers

Summary

BaRA is a framework for site-level web data collection combining bounded BFS traversal with history-based self-reflection, outperforming existing methods on link discovery and downloadable extraction.

arXiv:2607.00007v1 Announce Type: cross Abstract: Large language model (LLM)-based web agents reduce manual scripting for web data collection, yet on live websites, they often miss relevant pages, return incomplete multimodal outputs, or return media URLs that are not directly downloadable. We present BFS-and-Reflection Agent (BaRA), a framework for site-level collection under a fixed interaction budget. The framework combines bounded breadth-first search (BFS) traversal with history-based self-reflection. We evaluate BaRA on 50 synthetic websites with ground-truth reference sets. We additionally test on three public websites with cluttered or dynamic layouts. BaRA outperforms Pure LLM, SeeAct-Vision, and Browser-use on link discovery and downloadable multimodal extraction, with the largest gains in download-valid image and video recovery. Our code is available at https://github.com/MLAI-Yonsei/BaRA-Agent.
Original Article
View Cached Full Text

Cached at: 07/02/26, 05:41 AM

# BaRA: BFS-and-Reflection Web Data Collection Agent
Source: [https://arxiv.org/abs/2607.00007](https://arxiv.org/abs/2607.00007)
[View PDF](https://arxiv.org/pdf/2607.00007)

> Abstract:Large language model \(LLM\)\-based web agents reduce manual scripting for web data collection, yet on live websites, they often miss relevant pages, return incomplete multimodal outputs, or return media URLs that are not directly downloadable\. We present BFS\-and\-Reflection Agent \(BaRA\), a framework for site\-level collection under a fixed interaction budget\. The framework combines bounded breadth\-first search \(BFS\) traversal with history\-based self\-reflection\. We evaluate BaRA on 50 synthetic websites with ground\-truth reference sets\. We additionally test on three public websites with cluttered or dynamic layouts\. BaRA outperforms Pure LLM, SeeAct\-Vision, and Browser\-use on link discovery and downloadable multimodal extraction, with the largest gains in download\-valid image and video recovery\. Our code is available at[this https URL](https://github.com/MLAI-Yonsei/BaRA-Agent)\.

## Submission history

From: Soojeong Lee \[[view email](https://arxiv.org/show-email/8dae5879/2607.00007)\] **\[v1\]**Sat, 2 May 2026 08:09:10 UTC \(1,151 KB\)

Similar Articles

ARBOR: Online Process Rewards via a Reusable Rubric Buffer for Search Agents

arXiv cs.CL

ARBOR introduces a reusable rubric buffer to provide online process rewards for LLM-based search agents, improving training efficiency when outcome-only rewards are insufficient. It outperforms GRPO and DAPO on multi-hop QA benchmarks, converting up to 42% of zero-gradient training groups into informative ones.

Region4Web: Rethinking Observation Space Granularity for Web Agents

arXiv cs.CL

This paper introduces Region4Web, a framework that improves web agent performance by organizing observation spaces into functional regions rather than individual elements. It demonstrates that this approach reduces observation length and increases task success rates on the WebArena benchmark.