Source-free universal domain adaptation (SF-UniDA) adapts a pre-trained source model to an unlabeled target domain under both covariate and label shifts, without access to source data. However, existing SF-UniDA methods rely on inefficient techniques such as threshold tuning and clustering. Foundation models (FMs), known for their generalization and zero-shot capabilities, remain underexplored in SF-UniDA. In this paper, we propose a framework that leverages foundation models (LFM) for SF-UniDA. We use a vision-language model (VLM) to compute similarities between target samples and text labels, including those for unknown classes generated by prompting a large language model. The label shift type is determined by analyzing the coefficient of variation of a similarity-based sample-level score. Unknown samples are identified using a binary Gaussian mixture model fitted to another similarity-based metric. Under a consensus strategy, the pseudo-labels generated by the VLM are refined by the target model initialized with the pre-trained source model, integrating knowledge from both the source domain and foundation models. Finally, these refined pseudo-labels are used to train the target model. Extensive experiments across all possible label shifts and multiple benchmarks demonstrate the effectiveness and superiority of our proposed LFM framework. Our code is available at https://github.com/iamjingli/LFM.
Crystal Structure Prediction (CSP), the task of determining stable atomic arrangements from chemical composition alone, remains a central challenge in computational materials science with direct implications for accelerating materials discovery. While recent diffusion-based generative models achieve impressive results by conditioning on space-group symmetry information, this paradigm exposes three fundamental challenges: (1) unreliable symmetry inference from composition, (2) reliance on symmetry-only priors without comprehensive structural geometric guidance, and (3) prior-induced distribution shift caused by inaccurate or overly strong constraints. To address these challenges, we propose TGD-CSP, a three-stage generative framework that: (1) learns a cross-modal embedding space to retrieve structurally relevant templates directly from composition, thereby providing reliable symmetry priors; (2) guides diffusion-based generation via score-based conditioning that explicitly incorporates comprehensive geometric information from retrieved templates; and (3) fine-tunes the generative policy via reinforcement learning with an energy-based reward to alleviate prior-induced distribution shift and mitigate biased generation. % Evidence: Concrete results TGD-CSP achieves match rates of 74.56% and 75.25% on Perov-5 and MP-20, respectively, including a 42% relative improvement over state-of-the-art methods on Perov-5. It reduces RMSE to 0.0259 and 0.0258, respectively, while yielding structures with formation energies closely matching those of ground-truth structures. Our results demonstrate that TGD-CSP enables reliable generation from composition alone, significantly improving the practicality of generative models for real materials discovery.
Lu Yang, Tiantian Xu, X. Liu et al.· Proceedings of the 32nd ACM...· 0 citations