Reasoning-Driven Multimodal LLMs for Domain Generalization

Zhipeng Xu, Zilong Wang, Xinyang Jiang, Dongsheng Li, De Cheng, Nannan Wang

Paper code

• We propose RD-MLDG, a reasoning-driven multimodal LLM framework for domain generalization, which leverages class-relevant reasoning chains as domain-invariant supervision beyond visual features. We construct DomainBed-Reasoning to study reasoning under domain shift, and design MTCT to align reasoning supervision with direct classification and SARR to reduce reasoning-pattern mismatch through self-aligned reasoning refinement, enabling robust and interpretable out-of-domain generalization.

Prompt Disentanglement via Language Guidance and Representation Alignment for Domain Generalization

De Cheng, Zhipeng Xu, Xinyang Jiang, Dongsheng Li, Nannan Wang, Xinbo Gao

Paper

• We propose PADG for domain generalization, which combines LLM-guided prompt disentanglement with worst-case representation alignment. PADG uses CMD to learn domain-invariant and domain-specific visual features from disentangled textual guidance, designs WERA to simulate worst-case domain shifts and enforce representation alignment, and further adopts DSPL to integrate domain-invariant and domain-specific knowledge during inference.

[Symbiosis-Inspired Knowledge Distillation for Incremental Object Detection]

Mingyue Zeng, De Cheng, Zhipeng Xu, Huaijie Wang, Nannan Wang, Xinbo Gao

[Paper]

• We propose SIKD for incremental object detection, which leverages object symbiosis to preserve knowledge across old and new classes. SIKD designs SpSD to distill spatial dependencies from co-occurrence and occlusion regions, and SeSD to maintain old-class semantic topology through confidence-weighted prototype rank alignment, enabling a unified feature space that reduces old-new confusion and catastrophic forgetting.

Interference-Isolated Elastic Weight Consolidation and Knowledge Calibration for Incremental Object Detection

De Cheng, Mingyue Zeng, Zhipeng Xu, Di Xu, Nannan Wang, Xinbo Gao

Paper

• We propose IIKC for incremental object detection, which tackles catastrophic forgetting by explicitly modeling task interference between old and new classes. The core IKI-EWC module isolates conflicting knowledge caused by unlabeled objects being learned as background, and selectively regularizes model parameters to preserve useful old knowledge while suppressing interference. In addition, PKC calibrates old-class prototypes through a learnable projection layer and combines them with current-task features to reduce semantic drift in the classification head.

StPR: Spatiotemporal Preservation and Routing for Exemplar-Free Video Class-Incremental Learning

Huaijie Wang, De Cheng, Guozhang Li, Zhipeng Xu, Lingfeng He, Jie Li, Nannan Wang, Xinbo Gao

Paper

• We propose StPR, a unified exemplar-free framework for video class-incremental learning. StPR tackles catastrophic forgetting under spatiotemporal shifts by explicitly disentangling frame-shared semantics and temporal dynamics. Specifically, we design Frame-Shared Semantics Distillation (FSSD) to preserve stable semantic channels, and Temporal Decomposition-based Mixture-of-Experts (TD-MoE) to dynamically route temporal experts without task IDs or stored exemplars.

Adversarial Domain Prompt Tuning and Generation for Single Domain Generalization

Zhipeng Xu, De Cheng, Xinyang Jiang, Nannan Wang, Dongsheng Li, Xinbo Gao

Paper

• We propose PAPT for single domain generalization, which uses pre-trained text-to-image foundation models to synthesize diverse training domains from a single source domain. Instead of relying on hand-crafted textual styles, PAPT learns abstract category prompts and domain prompts to preserve class semantics while generating diverse domain shifts, and further designs a progressive adversarial training strategy to continuously expand the domain prompt pool.

Disentangled Prompt Representation for Domain Generalization

De Cheng, Zhipeng Xu, Xinyang Jiang, Nannan Wang, Dongsheng Li, Xinbo Gao

Paper

• We propose DPR for domain generalization, which leverages the easier disentanglement of language prompts to guide visual representation learning. Specifically, DPR uses LLM-generated domain-invariant and domain-specific descriptions to learn disentangled prompts, and designs Domain-Specific Prototype Learning (DSPL) to combine invariant prediction with relevant domain-specific knowledge for robust unseen-domain recognition.