Provenance-Aware Retrieval and Evaluator Consensus for Low-Resource Domain Adaptation of Large Language Models
Journal of Social Sciences, Humanities and Innovation · 07.10.2026 · 2 (2) · 235-248
DOI: 10.66238/jsshi.v2i2.59 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Social Sciences, Humanities and Innovation, online erschienen am 07.10.2026. Verfasst von Benjamin Johnson und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The adaptation of large language models to low-resource domains remains a formidable challenge in contemporary artificial intelligence research. Standard fine-tuning approaches often require extensive domain-specific corpora, which are inherently scarce in specialized fields such as rare disease diagnostics, indigenous language processing, and niche legal frameworks. Retrieval-augmented generation has emerged as a promising alternative; however, in low-resource settings, the quality of retrieved information is highly variable, and models are prone to hallucinating connections between disparate or unreliable source documents. This paper proposes a novel framework integrating provenance-aware retrieval with an evaluator consensus mechanism to address these limitations. By explicitly modeling the origin, credibility, and temporal relevance of retrieved documents, the provenance-aware module ensures that only highly reliable context is provided to the generation module. Concurrently, the evaluator consensus mechanism employs a multi-agent validation protocol, wherein independent evaluator agents assess the retrieved context and proposed responses, requiring a weighted consensus before final generation. This dual approach significantly mitigates hallucination and enhances factual accuracy in knowledge-starved domains. Extensive empirical analysis across three distinct low-resource datasets demonstrates that the proposed framework outperforms conventional adaptation methods in precision, recall, and factual consistency. The results indicate that strict provenance tracking combined with multi-agent consensus provides a robust, scalable solution for deploying large language models in specialized environments where data scarcity and reliability are critical concerns.
Zitieren
APA
Johnson, B., Reed, E., & Reyes, A. (2026). Provenance-Aware Retrieval and Evaluator Consensus for Low-Resource Domain Adaptation of Large Language Models. Journal of Social Sciences, Humanities and Innovation, 2(2), 235-248. https://doi.org/10.66238/jsshi.v2i2.59
Vancouver
Johnson B, Reed E, Reyes A. Provenance-Aware Retrieval and Evaluator Consensus for Low-Resource Domain Adaptation of Large Language Models. Journal of Social Sciences, Humanities and Innovation. 2026;2(2):235-248. doi:10.66238/jsshi.v2i2.59
BibTeX
@article{Johnson2026,
author = {Johnson, Benjamin and Reed, Elena and Reyes, Alyssa},
title = {Provenance-Aware Retrieval and Evaluator Consensus for Low-Resource Domain Adaptation of Large Language Models},
journal = {Journal of Social Sciences, Humanities and Innovation},
year = {2026},
volume = {2},
number = {2},
pages = {235--248},
doi = {10.66238/jsshi.v2i2.59}
}
EUVIMED