Which Negatives Matter? Ask Your Text Encoder: Adaptive Similarity Margins for Dense-Caption Retrieval
Researchers have developed a new method called HN-CLIP to improve dense-caption retrieval in AI systems. The current methods for this task often suffer from premature saturation due to the use of the InfoNCE objective and the presence of near-duplicate captions in benchmarks. HN-CLIP addresses these issues by using the text encoder's own geometry to construct adaptive similarity margins, which allows it to improve performance while training faster than other methods.
Researchers have developed a new method called HN-CLIP to improve dense-caption retrieval in AI systems. The current methods for this task often suffer from premature saturation due to the use of the InfoNCE objective and the presence of near-duplicate captions in benchmarks. HN-CLIP addresses these issues by using the text encoder's own geometry to construct adaptive similarity margins, which allows it to improve performance while training faster than other methods.
---
Why it matters: This matters because dense-caption retrieval is a crucial task in AI, particularly for applications like image captioning and visual question answering. Improving this task can have significant impacts on the accuracy and efficiency of these systems.
Source: https://arxiv.org/abs/2608.18521
This article was originally published at: https://arxiv.org/abs/2608.18521