Xuzhao Shi

Multimodal sarcasm detection aims to identify conflicts betweenthe intended and literal meanings of image--text posts by combiningholistic semantics with sparse local cues. Local retrieval requires both correspondence weights and content features. We propose Role-Decoupled Cross-Modal Evidence Retrieval (RDCER) to assign separate representation sources to these two operations. RDCER uses a CLIP-anchored global branch for holistic prediction, while a local branch retrieves complementary token--patch evidence. Within local retrieval, CLIP-derived representations compute correspondence weights, whereas calibrated RoBERTa and ViT features supply the content aggregated by those weights. Inverse-Anchor Retrieval Enhancement (IARE) adjusts the query--key features using local--anchor dissimilarity. It adds anchor context to these features while leaving the calibrated value features unchanged. The local branch further aggregates bidirectional token--patch evidence into local logits that are combined with the global prediction through learnable additive fusion. Experiments on MMSD and MMSD2.0 show that RDCER achieves accuracies of 93.44\% and 86.26\%, respectively. These results support the proposed representation assignment within the evaluated global--local framework, where holistic semantics and retrieved token--patch content contribute complementary classification information.

Multimodal sarcasm detection (MSD) aims to identify sarcastic expressions by integrating multimodal and contextual cues to capture cross-modal semantic inconsistencies. However, existing studies face challenges: fine-grained sarcastic cues are implicit and dispersed, large generative models suffer from gradient vanishing in classification tasks, and cross-domain generalization remains limited. To address these limitations, we propose InterARM, an interpretable affective reasoning model that introduces a structured sarcasm reasoning paradigm. Specifically, we introduce a three-stage training strategy based on curriculum learning, consisting of 1) sarcasm classification learning, 2) structured sarcasm reasoning learning, and 3) step-selective hierarchical reward reinforcement learning. The model generates reasoning chains across modalities to analyze sentiment polarity and semantic conflicts. We also construct the MSD-CoT dataset with 3,200 image-text pairs with 19,200 human-annotated reasoning steps. Experiments show that InterARM achieves superior performance on the in-domain and out-of-domain datasets, outperforming larger MLLMs such as Qwen2.5-VL-7B and GPT-5-mini while maintaining high interpretability and strong cross-domain generalization.