LLM Safety & Guardrails
Finding failure modes in safety systems and building guardrails that generalize beyond their training distributions.
Md Asiful Islam
asifulislamarizonaedu
I am a Computer Science PhD candidate at the University of Arizona. I study how Large Language Models (LLMs) can be made safer, more explainable, and more efficient, with recent work on LLM guardrails, reasoning, and information retrieval.
Research
Research focus
I work across model safety, reasoning, and efficiency—especially on systems that can explain why they made a decision.
Reasoning & Explainability
Designing reasoning methods that make safety decisions reliable, inspectable, and grounded in evidence.
Efficient NLP Systems
Developing lightweight models and retrieval systems that deliver strong performance with practical compute costs.
Selected publications
Recent work
A Dual-Hypothesis Reasoning Framework for LLM Guardrails
A reasoning framework that evaluates both safe and unsafe interpretations before making a safety decision, paired with multi-component supervised fine-tuning.
A Lightweight Explainable Guardrail for Prompt Safety
A compact multi-task guardrail that jointly classifies prompt safety and extracts word-level explanations, outperforming substantially larger models.
A Framework to Retrieve Relevant Laws for Will Execution
A hybrid retrieval pipeline combining lexical search, dense retrieval, reciprocal rank fusion, and LLM-based law selection.
Towards Realistic Few-Shot Relation Extraction: A New Meta Dataset and Evaluation
A realistic relation-extraction meta-dataset with unseen relations, limited supervision, non-uniform classes, and prevalent NOTA examples.
Best of Both Worlds: A Pliable and Generalizable Neuro-Symbolic Approach for Relation Classification
A modular neuro-symbolic classifier that combines transparent syntactic rules with a transformer-based semantic rule matcher.
VisText-Mosquito: A Unified Multimodal Dataset for Visual Detection, Segmentation, and Textual Explanation on Mosquito Breeding Sites
A multimodal dataset and benchmark for breeding-site detection, water-surface segmentation, and natural-language explanation.
Career
Work Experience
Jan 2026 — Jul 2026
Graduate Research Assistant
University of Arizona · NASA JPL OML Copilot
Developed an LLM agent that helps users write Ontological Modeling Language from natural-language specifications.
May 2025 — Aug 2025
Graduate Research Assistant
University of Arizona · Legal information retrieval
Designed and implemented a hybrid RAG framework for retrieving relevant laws from complete statutory corpora.
May 2024 — Aug 2024
Graduate Research Assistant
University of Arizona · Geospatial text analysis
Developed keyword extraction, geospatial text visualization, and LLM-based topic-title generation methods.
Jan 2018 — May 2022
Lecturer
American International University-Bangladesh
Taught artificial intelligence, algorithms, data structures, and object-oriented programming in C++.