Official Review · ICLR 2026

In this paper, the authors propose two components for utilizing unlabeled data during semi-supervised domain generalization training. Specifically, De-False-Correlation Adapter (DFC-Adapter) tries to reduce false correlations from domain-specific biases and pre-trained knowledge to refine visual features. And Learnable Multi-granularity Text-guided Embedding Augmentation (LMTEA) synthesizes domain-perturbed augmented visual embeddings for consistency regularization. 1. This paper proposes two components, including DFC-Adapter and LMTEA, to adapt pre-trained VLMs to SSDG without overfitting. 2. The proposed methods improved the utilization rate and accuracy of unlabeled data a lot. 1. The details of Figure 5 are unclear. What do the different colors represent - domains or categories? Specifically, does each cluster correspond to samples from the same category across different domains, or samples from the same domain across different categories? 2. The manuscript requires more rigorous proofreading and editing. In particular: (a) Tables 6 and 8 appear to be identical, and Tables 7 and 9 are also identical. (b) Line 136 contains a typographical error: "underexplored.." (duplicate punctuation). 3. Given the strong pretrained knowledge of CLIP, it would be more convincing to include its zero-shot performance as a reference in the experimental comparison. For instance, on DomainNet with CLIP ViT-B/16, the reported zero-shot accuracy is around 57.8% [1,2], suggesting that the results in Table 2 may not be directly comparable. 4. Additionally, since the proposed method leverages LLMs to augment domain and attribute prompts, it would be appropriate to include zero-shot CLIP results with the same prompt augmentation as an additional baseline. 5. The paper states that Equation (6) holds "based on the assumption that the predefined texts would be a well-matched pair for the image features." However, in Equation (8), the augmentation follows the same structure, but E(ta) may also not be well-matched to the target image features either. As a result, the augmented embeddings might not be optimal. [1] PromptStyler: Prompt-driven Style Generation for Source-free Domain Generalization.ICCV 23 [2] UMFC: Unsupervised Multi-Domain Feature Calibration for Vision-Language Models. NeurIPS 24 See Weakness part