Diffusion Integrated Gradients: Controllable Path Generation for Flexible Feature Attribution

Researchers have proposed Diffusion Integrated Gradients (DiffIG), a method that reframes how artificial intelligence models explain their decisions by generating attribution paths through a conditional generative process rather than relying on fixed trajectories. The technique, detailed in a paper posted to the arXiv preprint server on June 21, 2026, targets a known weakness in path-based attribution methods such as Integrated Gradients (IG) [1]. IG attributes a model’s prediction to input features by integrating gradients along a path from a baseline to the input, but the choice of that path heavily influences explanation quality. Existing approaches use fixed or hand-crafted paths that “often produce noisy or distorted attributions,” the authors write [2]. DiffIG reformulates path generation as a conditional generative modeling problem [3]. The method first trains a diffusion model to learn a distribution over paths generated from a Stick-Breaking Process, capturing diverse trajectories of varying form and complexity [4]. At inference time, it employs guided sampling to embed user-defined priors, steering path generation toward properties such as faithfulness and complexity [5]. This transforms attribution from a fixed procedure into a flexible, controllable generative process, the paper states [5]. The authors report that DiffIG quantitatively matches or outperforms existing path-based methods and produces explanations that are both perceptually aligned and faithful [2]. The work introduces what the researchers call “a new generative perspective for flexible, inference-time controllable Explainable Artificial Intelligence (XAI) methods” [1]. The paper was submitted by Soyeon Kim and revised two days later, on June 23, 2026 [1]. It appears under the Machine Learning category on arXiv, an open-access repository that hosts electronic preprints across physics, mathematics, computer science, and related fields [10]. arXiv, which began in 1991, passed the two-million-article milestone by the end of 2021 and now receives roughly 24,000 submissions per month [10]. Explainable AI remains a priority across the research community. Major laboratories such as Google DeepMind, a subsidiary of Alphabet Inc. formed through the 2023 merger of DeepMind and Google Brain, have advanced both foundational models and interpretability tools [6]. DeepMind’s portfolio includes the Gemini family of large language models and the AlphaFold protein-structure database, which released over 200 million predicted structures in 2022 [6]. The DiffIG paper is accessible through arXiv’s abstract page, which also surfaces experimental community tools under the arXivLabs framework [8]. arXivLabs, launched in 2020, allows third-party collaborators to build features such as citation explorers and code finders directly on the site while adhering to user-privacy commitments [8].