Institutional Repository of Vilnius University
Autonominė daugiagentė sistema CV ir darbo skelbimų semantiniam atitikimui ir sprendimų priėmimui /
Abstract
dc:descriptionThis thesis presents a multi-agent large language model system for semantic matching of resumes and job descriptions. Traditional applicant tracking solutions rely on keyword search or shallow sentence embedding similarity, missing context, multi-step reasoning and implicit requirements. Additionally, publicly available comparisons of LLM-based systems typically use a single model or a single evaluation reference, while dataset label reliability is rarely audited separately. An eight-agent specialised pipeline was designed, communicating through a shared blackboard context and implementing a reflection–revision loop, ESCO ontology-based contextual augmentation and two-pass calibration with a label memory (scenarios A, B and C). The system was evaluated using the public Resume-Screening-Dataset (5,000 pairs in the main experiment) and two independent references — the dataset's source labels and the verdicts of an independent LLM judge (DeepSeek-V3.2). A cross-model comparison (Qwen 7B) and a dataset augmentation experiment were carried out as additional studies. The agentic architecture surpassed both classical baselines — the Sentence-BERT embedding baseline and TF-IDF with a logistic regression classifier — against both references, yet the A→B→C scenario progression did not yield monotonic improvement. The main cause is dataset label noise: the independent audit found 31.8% of source labels to be incorrect, with 20.9% of cases identified as 'templated rejections'. Cross-model pairwise agreement reaches only 54.5%, and extending descriptions to a realistic length reduces F1. Both findings indicate that fair method comparison requires identical input form, and that a single reference point is not sufficient on its own.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Miežetis, Karolis,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308338050&prefLang=en_US