New AI Reads Job Ads and Sorts Skills Like an HR Expert

Researchers took on a task no automated LLM-based system had reportedly tackled before: extracting skills from free text in the form of (skill, level) pairs mapped to the Skills Framework for the Information Age (SFIA). That framework defines 147 professional skills across seven responsibility levels — capturing not just what skill is practiced, but at what responsibility level. The team tested five strategies: a lexical baseline, dense retrieval with LLM reranking, a zero-shot schema-constrained LLM, single-agent agentic RAG, and a three-agent retriever–matcher–verifier crew, evaluating them against expert-mapped European ICT role profiles. The results: retrieval-based matching identified the most skills, while generative strategies were markedly more precise. Only strategies that assigned the level as an explicit decision predicted it reliably, with similarity-based selection more than twice as inaccurate. And the three-agent crew doubled latency without improving accuracy — added agent roles don't automatically pay off for closed-taxonomy matching. The researchers also released the SFIA 9 corpus their pipeline was built on, offering what they describe as the first reproducible baseline for structured, level-aware skill extraction against SFIA.