Vision language model components differ in gender bias sensitivity
Mechanism-Level Evaluation for Vision-Language Models: Controlled Activation-Replacement Diagnosis of Gender Bias
Multimedia
Summary
Vision-language AI models can show gender bias, but it’s not clear which parts of the model cause that bias. The authors developed a way to check which internal layers of the model react most when gender-related changes are made. They found that language-processing parts of the models are often most sensitive to bias. This method helps reveal hidden biases inside the model’s structure that simpler tests might miss, but it does not by itself tell developers the best way to fix those biases.
What this means in practice
- •For machine learning engineers: Diagnose which specific parts of vision-language models are most sensitive to gender bias for better targeted fairness interventions.
- •For natural language processing teams: Combine mechanism-level evaluation with standard bias tests to improve understanding of bias sources across different model architectures.
Authors
Zhipeng Zhao, Wenxu Wang, Peishun Liu, Ruichun Tang
Abstract
Behavioral benchmarking reveals \emph{what} biases exist in vision-language models but not \emph{which internal components} are most sensitive to targeted intervention, precluding principled intervention. We argue for mechanism-level evaluation as a necessary complement, demonstrating causal mediation analysis as a diagnostic instrument for gender bias. We decompose gender-cue effects into controlled indirect effects attributable to specific-layer activations and direct effects through all other pathways, producing layer-by-layer mechanistic signatures. Across six models spanning three architectural families (LLaVA-1.5, LLaVA-NeXT, InstructBLIP at 7B/13B) and two 8B-scale architectures, three findings emerge: language-layer activations exhibit the greatest output sensitivity under controlled intervention, with the direct component often carrying the opposite sign; architectural choices redistribute layer-wise sensitivity to activation replacement; and counterfactual scores diverge from surface-level scores, exposing implicit associations. Systematic ablation validates internal consistency. An intervention experiment finds that the average indirect effect (AIE) and downstream intervention effectiveness are only weakly correlated (Pearson $r = 0.33$), and the layer with the second-largest AIE produces near-zero bias change---indicating that mechanistic diagnosis captures activation-replacement sensitivity but does not, by itself, identify optimal intervention targets. These results show mechanism-level evaluation captures architecture-specific sensitivity patterns that behavioral benchmarks cannot; pairing both should become standard NLP practice. Code: https://github.com/zhaozhipeng1997/CARD-GenderBias.