UGID: Unified Graph Isomorphism for Debiasing Large Language Models

2026-03-19

AI Summary

This work proposes a novel approach to mitigating social biases embedded in the internal representations of large language models, which are resistant to conventional output- or data-level debiasing techniques. By modeling the Transformer as a structured computation graph, the method introduces graph isomorphism principles to enforce structural invariance under counterfactual inputs. This invariance jointly constrains bias-sensitive regions at both the attention routing and hidden representation levels, thereby preventing bias propagation across model components. The framework integrates logit-space sensitivity constraints with an anchor-based semantic preservation mechanism to achieve unified, representation-level debiasing. Experiments demonstrate that the approach significantly reduces model bias and internal structural disparities under both in-distribution and out-of-distribution settings, while preserving model safety and general capabilities.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Large language models (LLMs) exhibit pronounced social biases. Output-level or data-optimization--based debiasing methods cannot fully resolve these biases, and many prior works have shown that biases are embedded in internal representations. We propose \underline{U}nified \underline{G}raph \underline{I}somorphism for \underline{D}ebiasing large language models (\textit{\textbf{UGID}}), an internal-representation--level debiasing framework for large language models that models the Transformer as a structured computational graph, where attention mechanisms define the routing edges of the graph and hidden states define the graph nodes. Specifically, debiasing is formulated as enforcing invariance of the graph structure across counterfactual inputs, with differences allowed only on sensitive attributes. \textit{\textbf{UGID}} jointly constrains attention routing and hidden representations in bias-sensitive regions, effectively preventing bias migration across architectural components. To achieve effective behavioral alignment without degrading general capabilities, we introduce a log-space constraint on sensitive logits and a selective anchor-based objective to preserve definitional semantics. Extensive experiments on large language models demonstrate that \textit{\textbf{UGID}} effectively reduces bias under both in-distribution and out-of-distribution settings, significantly reduces internal structural discrepancies, and preserves model safety and utility.

Problem

• social bias
• large language models
• internal representations
• debiasing
• graph isomorphism

Innovation

• Graph Isomorphism
• Internal Representation Debiasing
• Attention Routing
• Counterfactual Invariance
• Log-space Constraint