Reliable Large Language Models Through Ensemble Methods and Wisdom-of-the-Crowds Principles

I offer expertise in designing reliable Large Language Models (LLMs) using ensemble methods, uncertainty quantification, and the wisdom-of-the-crowds principles. My research focuses on building LLM systems that are robust, trustworthy, and suitable for deployment in high-stakes domains such as healthcare, law, education, and public administration.

This expertise is grounded in the CrowdLLM research project , which proposes a principled framework for combining multiple heterogeneous LLMs—ChatGPT, Llama, DeepSeek, etc.—to improve reliability beyond what any single model can achieve. The project develops theoretical and practical methods to: (i) formalize diversity, independence, and aggregation for probabilistic sequence predictors; (ii) quantify disagreement as epistemic uncertainty to detect hallucinations; (iii) align incompatible subword vocabularies; and (iv) incorporate semantic similarity between subwords to enable more effective LLM fusion.

I am interested in collaborating with MSCA fellows on:
• reliable generative AI
• LLM ensembles and output aggregation
• hallucination detection and epistemic uncertainty
• PAC-Bayesian and statistical learning theory for sequence models
• evaluation of LLM robustness on factual, reasoning, multilingual, or domain-specific benchmarks
• trustworthy AI for high-stakes fields (healthcare, justice, public sector)

Fellows with backgrounds in machine learning, NLP, computational linguistics, trustworthy AI, Bayesian modelling, or statistical ML will find strong alignment with this opportunity.