X-FACT avoids collapsing multimodal inputs into text-only representations. Instead, it encodes each modality natively, applies position and modality embeddings, and performs joint multimodal reasoning to produce veracity decisions, evidence grounding, and concise explanations.

The joint reasoning pipeline improves grounding, robustness, and interpretability by producing evidence-grounded veracity labels and concise human-readable explanations tied to specific regions or frames.

MAFC-7 is a seven-role multi-agent architecture that separates retrieval, verification, opposition, and judging while elevating two safety-critical roles, Timekeeper (temporal validity) and Provenance/Forensics (source authenticity, trust, and poison detection), as blocking checks. It couples provenance-aware retrieval and poison-aware debate policies with span-faithful rationales to improve robustness, abstention-aware coverage, and explanation fidelity in open-world fact-checking.

TL;DR: This paper investigates three key dimensions of fact-checking: the incorporation of multiple viewpoints, the utilization of diverse sources, and the analysis of disagreement among them.

Abstract: The rapid spread of misinformation across digital platforms poses significant societal risks. Yet most automated fact-checking systems depend on a single knowledge source and prioritize only supporting evidence without exposing disagreement among sources, limiting coverage and transparency. To address these limitations, we present a complete system for open-domain fact verification (ODFV) that leverages large language models (LLMs), multi-perspective evidence retrieval, and cross-source disagreement analysis. Our approach introduces a novel retrieval strategy that collects evidence for both the original and the negated forms of a claim, enabling the system to capture supporting and contradicting information from diverse sources (Wikipedia, PubMed, Google). These evidence sets are filtered, deduplicated, and aggregated across sources to form a unified and enriched knowledge base that better reflects the complexity of real-world information. This aggregated evidence is then used for veracity classification using LLMs. We further enhance interpretability by analyzing model confidence scores to quantify and visualize inter-source disagreement. Through extensive evaluation on four benchmark datasets with five LLMs, we showed that knowledge aggregation not only improves claim classification performance but also reveals differences in source-specific reasoning. Our findings underscore the importance of embracing diversity, contradiction, and aggregation in evidence for building reliable and transparent fact-checking systems. Our full code is available on GitHub.

Large Language Models (LLMs) often degrade when prompts contain irrelevant or distracting content because they tend to attend to all tokens indiscriminately. Inspired by neural mechanisms of selective attention and inhibition, this project proposes a pre-reasoning pipeline that filters, organizes, and suppresses low-value context so model outputs remain invariant or minimally affected by distractors.

The framework decomposes context handling into five modular stages: (1) perceptual segmentation to produce coherent chunks; (2) a relevance gate that scores, reranks, and prunes with recall-friendly thresholds; (3) focused attention that orders evidence, groups related facts, and produces brief tie-in summaries; (4) inhibition that suppresses residual distractors via span-level rules, query-conditioned penalties, and (for open-weights models) attention masks; and (5) reasoning/generation over the cleaned context with lightweight verification and an on-the-fly invariance test.

Doctor of Philosophy (PhD) in Computer Science

George Mason University, Fairfax, VA

MS in Computer Science - Machine Learning Concentration

George Mason University, Fairfax, VA

Bachelor of Science (B.Sc.) in Computer Science & Engineering

University of Chittagong, Bangladesh

Graduate Teaching Assistant

George Mason University - Virginia, United States

Applied Scientist Intern

Amazon.com Services LLC - Buyer Abuse Prevention ML, Seattle, WA, United States

Graduate Research Assistant

George Mason University - Fairfax, Virginia, United States

Software Engineer

BJIT Group, Tokyo, Japan

Software Engineer Intern

BJIT Group, Tokyo, Japan