AI Adapts on the Fly to Sort Electronic Waste With Far Less Training Data

Writing in Scientific Reports, Aseni Senanayake, Mehrtash Harandi, Nilakshan Kunananthaseelan, Mehrdad Arashpour and colleagues describe a segmentation method that adapts large-scale vision foundation models at inference time, rather than requiring lengthy offline retraining on domain-specific data. Images of discarded electronics are heterogeneous and cluttered, with devices often overlapping, partially obscured and presented in unpredictable orientations. Labeled training data for such scenes is scarce, and conventional fine-tuning of large models is computationally expensive and slow.

Vision foundation models, which are pretrained on enormous image collections, generalize well across visual domains, but the authors note that applying them directly to fine-grained e-waste segmentation has remained limited. Their solution is test-time fine-tuning. Instead of permanently adjusting the model's parameters before deployment, the approach adapts the model during inference using a small, diverse and contextually relevant support set of reference images.

The key innovation is how that support set is assembled. The researchers introduce what they call active retrieval, a two-stage selection process that first preselects candidate images using cosine similarity in the model's feature space, then applies an informativeness acquisition function to pick query-specific, non-redundant examples. This ensures the support images are both relevant to the item being segmented and sufficiently varied to guide the adaptation without redundant computation.

Evaluated on a curated e-waste dataset, the method delivered its headline result: end-to-end processing time was reduced by approximately sixfold compared with parameter-efficient fine-tuning, while segmentation accuracy remained reliable despite using less domain-specific training data. The findings are model-level results on a curated dataset rather than demonstrations inside an operating recycling facility, and the authors frame the work as a scalable foundation for AI-assisted e-waste management. If such systems can be deployed at scale, they could improve resource recovery from discarded electronics, a goal the team links directly to resource conservation and environmental sustainability.

The study was published open access in Scientific Reports on 26 October 2026.

Subject of Research: Test-time fine-tuning of vision foundation models for data-efficient electronic waste image segmentation

Article Title: Test-time fine-tuning of large-scale vision foundation models for data-efficient electronic waste segmentation

Article References: Senanayake, A., Harandi, M., Kunananthaseelan, N., & Arashpour, M. (2026). Test-time fine-tuning of large-scale vision foundation models for data-efficient electronic waste segmentation. Scientific Reports. https://doi.org/10.1038/s41598-026-77031-5

Image Credits: AI Generated

DOI: 10.1038/s41598-026-77031-5

Keywords: electronic waste, deep learning, computer vision, test-time fine-tuning, vision foundation models, image segmentation, active retrieval, recycling, resource recovery, artificial intelligence, Test-time, fine-tuning