Abstract: Real-world visual recognition tasks face dual challenges during training and testing:training data usually exhibit a pronounced long-tailed distribution,which causes models to be biased toward head classes;meanwhile,in test-agnostic scenarios,the class distribution at test time is typically unknown and may change dynamically,making it difficult for traditional methods that rely on test priors or static aggregation to maintain stable performance.Existing multi-expert methods still suffer from coarse-grained expert specialization,excessive test-time adaptation parameters,and low routing efficiency for tail classes.To address these issues,this paper proposes a test-time adaptive framework with dynamic collaboration for long-tailed visual recognition,termed Dynamic Collaborative Framework(DCF).The proposed method adopts a shared backbone network with multiple micro-expert heads to improve specialization across different class regions through fine-grained expert modeling;meanwhile,it designs a group-wise training objective with prior correction and incorporates prediction-level diversity regularization as well as classifier-head orthogonality regularization to promote dynamically collaborative complementary representations among experts.At test time,it further introduces a lightweight test-time adaptation mechanism that optimizes only group-level aggregation weights,enabling dynamic adaptation to unknown test distributions without updating the backbone network or expert parameters;in addition,a fast aggregation variant based on group prototypes is developed to balance inference efficiency.Experiments on multiple mainstream long-tailed recognition benchmark datasets show that the proposed method achieves or surpasses existing state-of-the-art methods under most test distributions and exhibits stronger robustness,particularly in tail-biased scenarios.Ablation studies further verify the effectiveness of the expert collaboration design,dual regularization,and test-time self-supervised optimization.

Key words: Long-tailed recognition, Test-agnostic, Dynamic collaboration, Test-time adaptation, Micro-experts