BD-Merging: Bias-Aware Dynamic Model Merging with Evidence-Guided Contrastive Learning

Researchers introduce BD-Merging, a new AI framework that improves model merging for multi-task learning by addressing bias and distribution shift issues. The method uses uncertainty modeling and contrastive learning to create more reliable AI systems that can better handle real-world data variations.

• →BD-Merging addresses reliability issues in model merging when test data differs from training data distributions.
• →The framework introduces evidential heads to model uncertainty across unified label spaces in multi-task scenarios.
• →An Adjacency Discrepancy Score quantifies alignment among neighboring samples to guide learning decisions.
• →Discrepancy-aware contrastive learning aligns consistent samples while separating conflicting ones for better robustness.
• →Extensive experiments show BD-Merging outperforms existing model merging baselines in effectiveness and robustness.