Pratyashi Nath

2+ years of work experience

Jan 20258 moSep 2025

Silchar, Assam, India

This research aims to investigate whether hallucinations in Large Language Models (LLMs) can be detected using only the model’s internal confidence signals, without relying on external fact checking systems. The study involves generating responses to factual question answering benchmarks such as TruthfulQA or Natural Questions using an open-source LLM (e.g., LLaMA or Mistral), while extracting internal signals including token probability distributions, entropy, log-likelihood scores, and attention patterns during inference. These signals will be analyzed to determine whether statistical differences exist between factually correct and hallucinated responses. By labeling outputs based on ground truth comparison and conducting correlation and classification analysis, the research seeks to develop a lightweight hallucination scoring mechanism. The final objective is to evaluate whether internal uncertainty measures can reliably predict factual inaccuracies, thereby improving the reliability and trustworthiness of generative AI systems without increasing computational overhead.