BaRA agent boosts web data collection with BFS and self-reflection

Large language model (LLM)-based web agents promise to replace manual scripting for scraping, but in practice they often miss relevant pages, return incomplete multimodal outputs, or serve media URLs that can't be directly downloaded. A team of six researchers (Lee, Lee, Cho, Kim, Moon, Song) from an undisclosed institution (likely KAIST or similar, based on name patterns) proposes BaRA (BFS-and-Reflection Agent) to solve these issues. The framework marries bounded breadth-first search (BFS) traversal—a systematic way to explore pages—with history-based self-reflection, allowing the agent to learn from past mistakes and avoid dead ends. The work is available on arXiv (2607.00007) with accompanying code on GitHub.

BaRA was rigorously evaluated against three baselines (Pure LLM, SeeAct-Vision, and Browser-use) on 50 synthetic websites with known ground-truth reference sets, plus three real-world public websites with cluttered or dynamic layouts. The agent achieved the highest performance in link discovery and downloadable multimodal extraction, with the largest improvements in download-valid image and video recovery. The fixed interaction budget design makes it practical for real-world deployment, avoiding runaway costs while still achieving comprehensive coverage. For professionals automating data pipelines, BaRA offers a more reliable way to collect rich media from complex websites without manual intervention.