The core insight behind Cortex is that career guidance is really two coupled problems. The first is semantic intent understanding: translating free-form, natural language aspirations into formal occupational categories. The second is pedagogical sequencing: given a target occupation and a learner’s current skills, deciding which competencies must be acquired, in what order, and which skills depend on which. The authors argue that existing systems handle neither well. Keyword-matching recommendation engines miss the latent meaning behind a user’s words, and even when a destination is correctly identified, most platforms list relevant courses without accounting for prerequisites, producing learning paths that are tempting to click through but pedagogically incoherent. Cortex addresses both problems inside a single retrieval-augmented architecture.

Technically, the pipeline begins with sentence embeddings. The system encodes a user’s career statement into a dense vector using a sentence-transformer model, in this case the widely used all-mpnet-base-v2 checkpoint, and compares it against embedded representations of occupational descriptions drawn from ESCO, the European Skills, Competences, Qualifications and Occupations database. ESCO is a multilingual taxonomy maintained by the European Commission that maps thousands of occupations to the skills and competences each requires. Because the matching happens in embedding space rather than through literal keyword overlap, a user writing about “interpreting big data” lands near the Data Scientist entry even though the exact phrase appears nowhere in the taxonomy. The authors note that this dense retrieval step is what gives Cortex its tolerance for paraphrase, metaphor and vague phrasing in user input.

Once an occupation is identified, the second stage takes over. Cortex consults a knowledge graph constructed from ESCO’s skill relationships, stored locally in an SQLite database, to determine which skills are essential, optional or optional-grouped for the chosen career, and how those skills depend on one another. This graph is the backbone of prerequisite-aware sequencing: if skill B requires skill A, the graph encodes that edge, and any generated learning path must respect it. The researchers describe this as a form of constrained generation. Rather than letting a large language model freely invent a curriculum, the system retrieves the relevant subgraph of skills and dependencies and feeds it, together with the user profile and target occupation, into the model as grounding context.

The generation stage itself is where retrieval-augmented generation, or RAG, earns its name. Cortex passes the retrieved, taxonomy-constrained skill set to a large language model through carefully engineered prompts, asking it to order the skills into coherent learning sessions, add scaffolding such as motivating context and progression notes, and produce structured JSON output. Two commercial models were used in the experiments, Google’s Gemini 2.5 Flash and OpenAI’s GPT-5.2, allowing the team to compare behavior across providers. Crucially, the constraint enforcement does not rely on prompt compliance alone: after generation, the system filters the output against the ESCO database programmatically, discarding any skill the language model hallucinated that does not exist in the taxonomy. The paper includes a table of out-of-taxonomy skills produced by an unconstrained baseline, illustrating exactly the kind of drift this post-generation filtering eliminates.

The evaluation pipeline combines automated and statistical methods. The authors assessed whether generated paths were prerequisite-consistent, whether the skills belonged to the correct taxonomy, and whether the sequencing respected the dependency graph, using end-to-end test scenarios executed against the live system. Statistical comparisons between configurations employed standard paired tests, including McNemar’s test for correlated proportions and Wilcoxon signed-rank procedures, with effect sizes and multiple-comparison corrections drawn from the classical literature on statistical power and sequential testing. Bootstrapping was used for confidence estimation. Experimental results reported in the paper indicate that combining an LLM with a constrained RAG framework and knowledge-graph-based dependency modeling yields paths that are reliably prerequisite-aware across diverse career goals, whereas an unconstrained LLM left to its own devices frequently invents skills and violates ordering constraints.

The authors position Cortex within a rapidly growing literature on AI in education and career recommendation. Prior work includes knowledge-aware autoencoders for explainable recommendations, learning path recommendation based on multidimensional knowledge graphs, dense passage retrieval for open-domain question answering, and, more recently, GraphRAG-induced dual knowledge structure graphs for personalized learning path recommendation presented at AAAI 2026. Cortex’s stated scientific contribution is the end-to-end connection of these threads: semantic intent understanding through embeddings, prerequisite-aware sequencing over an official occupational taxonomy, and taxonomy-constrained LLM generation with programmatic validation. Where earlier systems typically solved one stage and assumed the rest, Cortex binds them into a single reproducible pipeline whose datasets and source code are publicly available on GitHub.

Why does this matter beyond the machine learning community? Labor economists and policy groups, including the authors of the World Economic Forum’s Future of Jobs report, have emphasized that reskilling at scale is becoming a structural necessity as automation and AI reshape job content. A system that can translate a vague aspiration into a concrete, ordered plan of skills, grounded in an authoritative multilingual taxonomy and validated against explicit dependencies, addresses one of the practical bottlenecks in lifelong learning: not finding content, but knowing what to learn first. The scaffolding produced alongside each path, in which the language model explains why a skill appears where it does, also speaks to a long-standing demand for explainability in recommender systems, echoing earlier work on knowledge-aware recommendation that made suggestions inspectable rather than opaque.

Limitations and open questions remain visible in the design. Cortex depends on the quality and coverage of ESCO, which, while extensive, reflects a particular European framing of occupations and competences; skills central to emerging roles may lag the taxonomy. The system also inherits the general fragilities of large language models, mitigated but not eliminated by the programmatic post-filtering. And while the evaluation demonstrates prerequisite consistency and taxonomy compliance, long-term studies of whether learners who follow Cortex-generated paths actually achieve their career outcomes more often would require longitudinal data the current paper does not include. The authors have nonetheless made the implementation, experimental scripts and evaluation pipeline openly accessible, and the architecture is deliberately modular: swap the embedding model, the taxonomy or the LLM, and the pipeline still functions. That modularity, as much as any single benchmark number, may determine whether frameworks like Cortex become the default plumbing behind the next generation of career guidance platforms.

Subject of Research: Retrieval-augmented generation of prerequisite-aware, career-aligned learning paths using large language models and the ESCO skills taxonomy

Article Title: Cortex: A Retrieval-Augmented Framework for Career-Aligned Learning Paths

Article References: Ramachandran, V., Vedanta, M., & Khetarpaul, S. (2026). Cortex: A Retrieval-Augmented Framework for Career-Aligned Learning Paths. Machine Learning, 115(10), Article 227. https://doi.org/10.1007/s10994-026-07160-5

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07160-5

Keywords: retrieval-augmented generation, large language models, learning path generation, knowledge graphs, career recommendation, sentence embeddings, ESCO, skill dependencies, personalized upskilling, AI in education, semantic matching, Shiv Nadar University