Disambiguation on Long-Tail Entities

Improve and characterize disambiguation for rare, long-tail entities in reasoning-native multimodal entity-linking models, which can retrieve the correct entity but still reject it in favor of an incorrect candidate.

Background

The proposed 8B-Think+Embed system combines iterative Wikipedia retrieval with reasoning over textual and visual context, substantially improving performance on rare entities. Nevertheless, the error analysis reports that some examples contain the correct entity in the retrieved evidence while the model ultimately rejects it. This indicates that retrieval alone does not resolve the candidate-selection problem, particularly for difficult long-tail entities and neighboring concepts.

The unresolved issue is specifically the model's ability to select and disambiguate among retrieved candidates after the correct entity has been surfaced, rather than merely improving retrieval recall.