This write-up is being prepared. Please check back soon, or email me at os9660@rit.edu.
MS Research · LLMs · Evaluation
Multi-objective evolutionary search for toxicity control in large language models using context-aware metrics.
How can we reduce toxic generations from large language models without collapsing fluency, and how should toxicity be measured when context changes the meaning of a phrase?
I framed toxicity control as a multi-objective search problem. Context-aware metrics scored candidate interventions, and evolutionary search explored trade-offs between safety and utility instead of optimizing a single scalar score.
The study produced a structured experimental pipeline for comparing context-aware toxicity metrics under multi-objective search. It is the core research thread of my MS work at RIT and the foundation I am carrying into PhD study.
Results depend on the chosen toxicity proxies and evaluation corpora; metric misspecification can reward shallow refusals. Broader human evaluation and stronger baselines remain open work.