Publications & Presentations

IKnowFlow is the doctoral thesis framework connecting my work on attribution, interpretable retrieval, and evidence grounding. The framework treats trustworthiness as a pipeline property: retrieval should be explainable, evidence selection should be justified, and generated claims should remain traceable to source material.

The consortium submission positions REASONS, IMRNNs, and METEORA as complementary pieces of that pipeline, then outlines future work on negation-aware retrieval and provenance-aware verification through structured source graphs.

Venue: IJCAI-ECAI 2026 Doctoral Consortium, Bremen, Germany. Travel grant awarded.

METEORA reframes the retrieval stage of RAG as an evidence selection problem rather than a fixed top-k ranking problem. The method uses rationale-guided selection to identify evidence that is actually useful for answering the query, reducing dependence on opaque re-rankers and arbitrary context cutoffs.

Across scientific, legal, and finance datasets, METEORA improves answer accuracy by +33.34% while using 50% fewer evidence chunks, and raises adversarial defense F1 from 0.10 to 0.44.

Visit project website 

IMRNNs addresses a central limitation of dense retrieval: retrieved documents are scored through opaque embedding similarity, making it difficult to understand why a document matched a query. The method adds lightweight query and document adapters over a frozen dense retriever, allowing representations to be modulated dynamically at inference time.

The interpretability contribution comes from analyzing how embeddings move after modulation and projecting those changes back into semantic directions associated with the base encoder. This gives users a way to inspect which concepts were emphasized or suppressed during retrieval.

View on ACL Anthology 

This article frames retrieval as a key trust bottleneck in RAG. If the retriever selects weak or misleading evidence, the generator inherits that limitation. The paper argues for neurosymbolic retrieval methods that combine the flexibility of neural retrievers with explicit symbolic structures that humans can inspect, validate, and use for debugging.

It discusses mechanisms such as knowledge-modulated retrieval, knowledge-graph path enrichment, and process-knowledge-guided ranking. Together, these approaches show how retrieval can move beyond black-box similarity scoring toward evidence selection that exposes the reasoning context behind a retrieved document.

View on IEEE Xplore 

This Wiley book chapter examines how neurosymbolic AI can support Legal AI-TRISM: trustworthy, reliable, interpretable, and safe AI for the legal domain. The chapter motivates the need for stronger grounding and governance when LLMs are used as legal assistants, where unsupported claims, hallucinated citations, and opaque reasoning can have serious consequences.

The chapter discusses the symbiosis of LLMs and knowledge graphs for legal RAG. Neural models provide flexible language understanding, while symbolic legal knowledge, graph structure, and rule-aware reasoning provide a basis for reliability, traceability, and safer decision support.

View on Wiley Online Library 

This paper studies how LLMs should attach citations to generated answers. Generation-Time Citation asks the model to cite evidence while generating text, while Post-hoc Citation first generates the answer and then links claims back to retrieved evidence.

The key takeaway is that attribution should be treated as a retrieval-sensitive problem rather than only a generation-formatting problem. In high-stakes settings, stronger retrieval and careful post-hoc verification can improve coverage while preserving citation correctness.

View preprint 

This paper evaluates whether LLMs are not only answering questions correctly, but also producing explanations that remain consistent with their answers. It uses benchmark-based evaluation and automatic metrics to compare public and proprietary models on answer quality and explanation quality.

The results show that strong performance on answers does not necessarily mean reliable reasoning. This finding motivates more careful evaluation of LLMs in settings where users rely on explanations, not just final outputs.

View preprint 

This work studies the cybersecurity implications of LLM-generated code transformations. It introduces MetamorphASM to evaluate whether language models can produce assembly-level obfuscations, including transformations that alter control flow, insert dead code, or substitute registers.

The findings show that LLMs can generate obfuscated code patterns that complicate malware detection and static analysis. The paper highlights a dual-use risk: the same code-generation capabilities that help developers can also be misused to hide malicious behavior.

View paper 

This paper presents a machine learning pipeline for classifying student mental health signals from social media text. It combines sentiment analysis and supervised learning models to identify emotional states such as anxiety, sadness, and joy.

The broader goal is early identification of possible mental health concerns in academic environments. While the work is exploratory, it demonstrates how NLP and emotion classification can support proactive student well-being interventions when used responsibly and with appropriate safeguards.

View paper