Marwan Farag

I am conducting research in the fields of Large Language Model (LLM) Hallucination and Uncertainty Estimation. Hallucination remains a critical issue, as LLMs occasionally generate fluent but non-factual or inaccurate text. Uncertainty Estimation is an active research area that seeks to quantify the level of confidence an LLM has in its outputs.

The goal of my research is to quantify this uncertainty and leverage it to improve LLM performance. For instance, an LLM system could utilize uncertainty metrics to abstain from answering or query an external oracle, rather than generating plausible yet inaccurate information. Furthermore, these uncertainty estimates can be used to enhance reasoning LLMs for improved test-time scaling, ensuring that specific rationales receive higher attention based on their reliability.