Publications & Presentations

This work extends interpretable retrieval by introducing efficient embedding modulation techniques that produce token-level explanations while reducing computational overhead in dense retrieval. By modulating embeddings at different layers, IMRNNs achieve better interpretability without sacrificing retrieval performance, making them suitable for applications requiring both accuracy and transparency.

This work proposes a neurosymbolic retrieval approach for RAG that integrates symbolic structure from knowledge graphs with neural retrieval, aiming to make evidence selection more explicit and auditable in high-stakes settings. It introduces three complementary mechanisms: MAR (Knowledge Modulation Aligned Retrieval) to refine query representations using interpretable symbolic features, KG-Path RAG to enrich queries via graph traversals, and process knowledge-infused RAG to reorder retrieved content using validated procedural workflows. Early results on mental health risk assessment tasks suggest improvements in both transparency and overall retrieval quality.

View on IEEE Xplore 

In this work, we investigate two paradigms for citing sources in generated text: Generation-Time Citation (G-Cite) and Post-hoc Citation (P-Cite). G-Cite annotates evidence as text is produced, whereas P-Cite retrospectively links generated answers to retrieved documents. Through experiments on multiple attribution datasets we show that citation quality is largely governed by the quality of retrieval. P-Cite yields higher coverage and competitive correctness compared to G-Cite, demonstrating that post-hoc approaches can provide more complete attribution while remaining precise. Our analysis suggests adopting a retrieval-centric, P-Cite-first strategy in high-stakes domains, where transparency and completeness of citations are paramount.

This paper benchmarks the consistency and reasoning abilities of large language models using the Boolq dataset. We assess both public and proprietary models across metrics including BERT-score, BLEU and F-1 on generated explanations and answers. Results reveal that proprietary models outperform open-source alternatives; however, no model excels at both consistency and reasoning simultaneously. The findings underscore gaps in current evaluation and highlight the need for models that can both reason and maintain internal consistency across tasks.

Addressing a growing cybersecurity question, this work introduces MetamorphASM, a benchmark of 328,200 obfuscated assembly code snippets, to evaluate whether LLMs can generate obfuscated malware. We test models under various obfuscation techniques, including dead code insertion, register substitution and control-flow manipulation, and find that LLMs are capable of producing obfuscated code that challenges static analysis tools. The study emphasises the dual-use nature of LLMs and calls for developing safeguards to mitigate malicious applications.

This paper presents a sentiment-driven machine learning pipeline for identifying the mental state of engineering students from social media posts. By combining sentiment analysis with recurrent neural networks (RNNs), gated recurrent units (GRUs) and support vector machines (SVMs), the system classifies emotions into categories such as anxiety, joy and sadness. The approach seeks to facilitate early detection of mental health concerns and encourage proactive support within academic communities.