Diffusion Integrated Gradients: Controllable Path Generation for Flexible Feature Attribution

Researchers introduce Diffusion Integrated Gradients (DiffIG), a novel explainable AI method that uses diffusion models to generate optimized attribution paths instead of relying on fixed hand-crafted paths. The approach enables inference-time controllable feature attribution with improved explanation quality and perceptual alignment compared to existing path-based methods.

Explainability in machine learning has become increasingly critical as AI systems make consequential decisions across finance, healthcare, and other sectors. Traditional path-based attribution methods like Integrated Gradients provide strong theoretical foundations but suffer from a fundamental limitation: their explanation quality depends heavily on the path chosen from baseline to input, yet practitioners have lacked principled ways to optimize this choice. DiffIG addresses this gap by treating path generation as a learnable problem rather than a fixed design choice. The method trains a diffusion model to capture a distribution over high-quality attribution paths, then uses guided sampling to incorporate user preferences during inference. This generative approach represents a meaningful shift in how the XAI community thinks about feature attribution—moving from static interpretability to dynamic, context-aware explanations. The technical contribution bridges generative modeling and interpretability research, two increasingly important areas in AI development. For practitioners deploying AI systems, DiffIG offers tangible benefits: more reliable explanations reduce risks of misinterpreting model behavior and strengthen regulatory compliance in industries requiring explainability. The inference-time controllability also enables domain experts to steer explanations toward their specific analytical needs without retraining models. As machine learning systems face growing scrutiny from regulators and end-users demanding transparency, methods that improve explanation quality without sacrificing computational efficiency gain strategic value. The work suggests that future XAI developments will likely incorporate similar generative approaches, allowing practitioners to move beyond one-size-fits-all explanations toward more flexible, problem-specific interpretability solutions.

• →Diffusion Integrated Gradients reformulates attribution path generation as a conditional generative modeling problem rather than relying on fixed paths.
• →The method enables inference-time controllable explanations through guided sampling, allowing users to steer attributions based on domain knowledge.
• →DiffIG quantitatively matches or outperforms existing path-based attribution methods while producing more perceptually aligned explanations.
• →This generative perspective opens new possibilities for flexible, context-aware explainable AI systems across industries.
• →The approach addresses a fundamental limitation in current XAI methods by optimizing the critical path selection component.