Lavdim Halilaj

Research profile

Research summary

Lavdim Halilaj is a researcher at Robert Bosch (Germany) whose record, resolved from OpenAlex bibliographic data, traces an unusually wide arc across applied AI. The topic profile is led by Semantic Web and Ontologies (21 works), followed by Advanced Graph Neural Networks (13), Service-Oriented Architecture and Web Services (11), Domain Adaptation and Few-Shot Learning (11), and Data Quality and Management (10) — a spread consistent with a path from enterprise software and service computing into modern machine learning. The current field split is even, 0.5/0.5, between causal-inference and explainable ML on one side and NLP/multimodal information retrieval on the other. Activity is the standout signal at 93/100: 23 works in the last three years and 13 already recorded in 2026, the highest annual count since the record begins in 2014 (h-index 14, i10-index 21). Two 2026 arXiv preprints are in the Evolve catalog. xWhyL formalizes learning causal models from expert explanations through explanatory equivalence classes and a Tug-of-War theorem, though code remains pending and validation is synthetic-only. Less Is More shows a graph-free multimodal RAG pipeline built on multi-signal late fusion matching graph-based systems at far lower cost, with results Evolve judges credible and honestly reported. Both papers list Halilaj as a middle author, consistent with a collaboration score of 48/100 over a stable circle of about 10 recurring co-authors.

Research keywords

Highlights

• xWhyL: learning causal models from expert explanations

  The 2026 preprint introduces the explanatory equivalence class (EEC), a Tug-of-War theorem and the CIL method for causal interactive learning. Evolve calls the formalization promising but notes code is pending and validation is synthetic-only (repro readiness: medium).

• Graph-free multimodal RAG matches graph systems at lower cost

  Less Is More replaces graph pipelines with multi-signal late fusion — reciprocal rank fusion, multi-vector retrieval and VLM-enhanced queries — for document question answering. Evolve judges the results credible and honestly reported, while AutoQA is only partly human-verified (repro readiness: medium).

• From semantic-web engineering to causal and multimodal ML

  The topic record spans Semantic Web and Ontologies (21 works) and Service-Oriented Architecture (11 works) alongside Advanced Graph Neural Networks (13) and Few-Shot Learning (11), with a current 0.5/0.5 field split between causal/explainable ML and NLP/multimodal retrieval — an arc both 2026 preprints reflect.

Timeline

1. 2014First works appear in the OpenAlex record (1 work in 2014).
2. 2016Output peaks early at 12 works — the highest annual count before 2026.
3. 2019Output dips to a single work, the lowest point in the record.
4. 2021A sustained run begins: 7 works in 2021, followed by 8 in 2022 and 8 in 2023.
5. 202613 works already recorded, the highest annual count in the record; two arXiv preprints — xWhyL and a graph-free multimodal RAG paper — enter the Evolve catalog.

Papers in the Evolve catalog

• 2026-09-22

  xWhyL: Causal Interactive Learning

• 2026-09-18

  Less Is More: Graph-free Multimodal RAG via Multi-signal Late Fusion

Frequent co-authors