International Journal of Data Science and Analytics Информационные системы
LLM-based agents for generalized web automation and schema-guided data extraction: a survey
Аннотация
Large language model (LLM) agents can navigate websites, operate graphical interfaces, and use software tools, but task completion does not guarantee reliable structured data. This survey examines the stronger problem of generalized schema-guided extraction: discovering sources, traversing unfamiliar sites, populating typed records with field-level provenance, and assembling a validated dataset without site-specific redevelopment. We conducted a PRISMA-informed systematic mapping review with narrative synthesis covering from January 2017 to April 2026. The mapping corpus contains 805 unique studies, of which 454 are archival published or formally accepted; a historical 403-record synthesis-candidate set yielded 385 retained studies after publication-status and full-text reconciliation; during reviewer-driven scope re-adjudication, one already-mapped, publication-eligible KDD 2025 survey (Record 760) was moved from the supplementary landscape into the synthesis, giving a final qualitative synthesis of 386 studies. All 386 synthesis studies were read in full: 114 P0/P1 studies received deep critical analysis and 272 P2/P3 studies complete structured reading. The synthesis connects observation, grounding, planning, memory, learning, security, and evaluation to a six-stage source-to-dataset pipeline. We propose a multi-level evaluation framework and ground it in a controlled 108-run case study crossing three agent architectures, three LLM backends, four frozen commerce sites, and three repetitions. The strongest observed configuration reaches macro Field F1 0.9440, while a Playwright-MCP configuration reaches macro Field F1 0.7872 but task success only 0.1667, illustrating that extraction quality and task completion can diverge. The central conclusion is that progress in Web agents should be assessed by verifiable outputs, not navigation success alone.