<p>Anaphora resolution (AR) is the process of identifying the antecedent of a pronoun or other anaphoric expression, occurring within a given text. It is essential for accurate semantic interpretation and disambiguation of text. In this paper, we propose a novel hybrid model for Hindi AR task. The model combines rule-based and machine learning approaches for handling different types of pronominal anaphoric references. The rule-based approach is used to resolve reflexive and distributive anaphoric references while stacked ensemble architecture is used to resolve demonstrative and relative anaphoric references in the Hindi text. The combined module efficiently resolves inter-sentential and intra-sentential anaphoric references utilizing syntactic features, semantic-disambiguators such as named-entity and animacy agreement, linguistic features and statistical metrics. We also investigate the efficacy of Instruction-tuned Large Language Models (LLMs) in conjunction with vanilla prompt engineering for AR in the Hindi text. By employing LLMs, we aim to overcome the probable antecedent’s detection limitation of traditional rule-based and corpus-based methods, thereby automating the process of identifying noun referents (or potential mentions) for an encountered anaphoric expressions. We used the CFILT Bombay Hindi Tourism dataset and publicly available Hindi Dependency Tree Bank dataset (HDTB) to develop, train and evaluate our proposed approach. Through fivefold cross-validation, our hybrid model achieved an accuracy of 77.205% on Hindi Tourism data. The proposed model significantly outperformed the state-of-the-art accuracy on the HDTB news dataset, achieving a remarkable 73.503% accuracy. Our experiments and results allude towards the effectiveness of the proposed approach, indicating a wider potential in the development of an automatic AR system for low resource languages.</p>