Comparative Analysis of Local Large Language Models for Ranking Higher Education Programmes Based on Applicant Digital Profiles

Choosing among closely related higher-education programmes requires interpretation of heterogeneous applicant data while preserving data confidentiality. This study compares ten locally executable large language model (LLM) configurations as semantic rankers, contextualises their performance against a hybrid term frequency–inverse document frequency (TF–IDF) cosine baseline, and evaluates robustness when relevant competencies are expressed indirectly. The main benchmark comprised 50 anonymised applicant profiles and 10 degree programmes; an additional processed-profile robustness set comprising 10 profiles was used to reduce direct lexical overlap with the programme catalogue. Rankings were evaluated using Accuracy@1, Accuracy@3, normalised discounted cumulative gain at 5 (NDCG@5) and mean reciprocal rank at 5 (MRR@5), while structured-output validity and inference time were assessed for the LLMs. Among the local LLMs, gpt-oss-20b-MXFP4 achieved the highest Accuracy@1 (0.76), whereas gemma-4-E4B-it-Q8_0 achieved the highest Accuracy@3 (0.92), and Ministral-3-14B-Reasoning-2512-Q4_K_M provided a favourable quality–latency balance. On the main benchmark, TF–IDF achieved Accuracy@1 = 0.88 and NDCG@5 = 0.9557, exceeding all LLM configurations and demonstrating a strong lexical signal. On the processed-profile robustness set, the Accuracy@1 of the TF–IDF baseline decreased to 0.60, whereas gemma achieved 0.90. The results provide initial evidence that several local LLM configurations can reproduce observed programme-selection patterns in a limited pilot benchmark. However, the findings should be interpreted as task-specific model-comparison results rather than as full validation of an autonomous career guidance system.