Recent advances in Multimodal Entity Linking (MEL) exploit textual and visual information to disambiguate mentions and align them with entities in a knowledge base. Existing methods typically design separate and complex network modules for each type of interaction among multi-granular and multimodal features, while lacking explicit modeling of the joint dependencies among these features. Moreover, most approaches rely on unidirectional retrieval-based matching and lack knowledge-driven verification, leading to unreliable disambiguation in weak-context scenarios. To address these challenges, we propose a novel two-stage MEL framework termed ThinkLinker. First, we introduce a low-rank fusion mechanism to model the joint dependencies among multi-granular and multimodal features, enabling comprehensive and explicit interactions while learning task-relevant discriminative information for candidate ranking in a lower-dimensional space. Subsequently, we develop a bidirectional retrieval-verification paradigm, where the ranked candidate entities guide an LLM-based multi-turn, dialogue-style verification process to generate mention-specific contextual augmentation. The augmented context is then adaptively fused with the original representation to further refine the linking model. Experimental results on public benchmark datasets demonstrate that the proposed ThinkLinker outperforms all state-of-the-art baselines. The code is publicly available at https://github.com/zhouyuanyu/ThinkLinker.

Multimodal Entity Linking (MEL) aims to link mentions in a given context to their corresponding entities in a knowledge base by leveraging both textual and visual information. Existing MEL methods still underexploit visual information in real-world scenarios. Most approaches directly encode raw image-text inputs or apply uniform enhancement, while lacking adaptive modeling of heterogeneous mention-relevant evidence. Moreover, they typically rely on single-layer visual representations or simple cross-modal interactions, failing to capture hierarchical visual semantics required for fine-grained disambiguation. To address these limitations, we propose a novel MEL framework termed EVEMEL. First, we introduce an adaptive evidence augmentation mechanism, which analyzes the visual condition of each sample and strengthens mention-relevant discriminative clues, providing more reliable contextual evidence for entity disambiguation. Subsequently, we develop a cross-modal modeling mechanism that treats multi-layer visual representations as different semantic granularities and dynamically fuses them through token-level routing, enabling cross-modal matching. In addition, a bidirectional gated intra-modal mechanism is designed to capture multi-granularity textual and visual matching signals. Experimental results on public benchmark datasets demonstrate that the proposed EVEMEL outperforms all state-of-the-art baselines. The code is publicly available at https://github.com/hanwang86/EVEMEL.