Patanjali Bhamidipati supervised by Dr. Radhika Mamidi received his Master of Science by Research – Dual Degree in Computational Linguistics (CLD). Here’s a summary of his Hallucination Detection in Large Language Models: A Zero-Shot Inference Framework and a Multilingual Scientific Benchmark

Large Language Models (LLMs) are highly fluent, but this fluency can mask a critical failure mode: hallucinations, i.e., outputs that are well formed yet factually incorrect, semantically unfaithful, or insufficiently grounded in available context. This thesis characterizes hallucination as a context-dependent grounding failure and investigates how it manifests across tasks, domains, and languages. In controlled NLG settings, we formalize hallucination detection as a task-aware inference problem. In definition modeling, hallucinations arise when generated definitions do not entail the intended meaning; in machine translation and paraphrase generation, they correspond to violations of semantic equivalence between the source and the output. Building on this formulation, we propose a zero-shot, computationally efficient detection framework based on pretrained Natural Language Inference (NLI) models that operates in both model-aware and model-agnostic settings. Experiments on the SHROOM benchmark show that this approach is competitive while avoiding opaque, expensive LLM-as-a-judge evaluation. We then demonstrate that inference-based signals are often insufficient in real-world scientific scenarios with weaker grounding and long contexts. To study this regime systematically, we introduce CAP (Confabulations from ACL Publications), a nine-language dataset for scientific hallucination detection consisting of publication-grounded questions, multiple LLM-generated answers, and expert annotations that explicitly separate factuality from fluency. Benchmarking diverse reference-based and referencefree detectors on CAP reveals substantial performance drops across domains and languages, especially in low-resource settings, highlighting the need for more robust grounding-aware methods for trustworthy generation. 

June 2026