UGID: Graph Isomorphism Debiasing Addresses LLM Social Bias at Structural Level
arXiv 2603.19144·medium signal
UGID (arXiv:2603.19144) applies unified graph isomorphism to identify and remove structural bias patterns in LLMs rather than patching outputs or resampling training data. The approach targets the internal computation structure where biases are encoded, claiming output-level and data-optimization methods cannot fully resolve social biases. Represents a mechanistically distinct debiasing strategy compared to RLHF or prompt-engineering approaches.