Research map / Vision-language and multimodal / Cross-modal learning
FALCON: False-Negative Aware Learning of Contrastive Negatives in Vision-Language Alignment
Myunsoo Kim, Seongwoong Shim, Byung-Jun Lee
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FALCON learned negative scheduler instead of GRIT-VLP fixed hardest-negative sampling | ALBEF · MSCOCO IRTR text retrieval | R@1: 60.60 -> 62.28 | Improved |
| add FALCON instead of MAFA false-negative relabeling | ALBEF · MSCOCO IRTR image retrieval | R@1: 44.77 -> 46.18 | Improved |
| use only easiest negatives (q=0.0) instead of hard ones | ALBEF · MSCOCO IRTR text retrieval | R@1: 62.28 -> 18.92 | Hurt |
| add FALCON to BLIP-2 | BLIP-2 · OKVQA test | accuracy: 17.94 -> 20.96 | Improved |
| add GRIT-VLP hard negatives to BLIP-2 | BLIP-2 · VQA2 val | accuracy: 42.46 -> 39.91 | Hurt |
| add FALCON to BLIP-2 | BLIP-2 · COCO Stage-1 text retrieval | R@1: 75.22 -> 75.56 | Improved |
| add FALCON to SigLIP-2 | SigLIP-2 · MSCOCO text retrieval | R@1: 69.96 -> 72.96 | Improved |
| add FALCON to SigLIP-2 | SigLIP-2 · MSCOCO image retrieval | R@1: 54.21 -> 54.15 | No clear effect |
| increase search space size from 480 to 5664 (ablation) | FALCON · MSCOCO IRTR text retrieval | R@1: 58.48 -> 61.72 | Improved |
| increase search space size from 5664 to 28320 (ablation) | FALCON · MSCOCO IRTR text retrieval | R@1: 61.72 -> 61.94 | No clear effect |
| use MLM alone as scheduler proxy instead of ITC+ITM (ablation) | FALCON · MSCOCO IRTR text retrieval | R@1: 57.64 -> 61.72 | Improved |
| use instance-level instead of batch-level scheduling (ablation) | FALCON · MSCOCO IRTR text retrieval | R@1: 58.78 -> 61.72 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- MAFA — MAFA: Managing False Negatives for Vision-Language Pre-training (CVPR 2024) · better on 6 of 6 shared results
- GRIT-VLP — GRIT-VLP: Grouped Mini-Batch Sampling for Efficient Vision and Language Pre-training (ECCV 2022) · better on 3 of 3 shared results
- BLIP-2 — BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation (ICML 2022) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MAFA: Managing False Negatives for Vision-Language Pre-training (CVPR 2024 · similarity 0.90)
- Enhancing Fine-Grained Vision-Language Pretraining with Negative Augmented Samples (AAAI 2025 · similarity 0.89)
- SaCo Loss: Sample-wise Affinity Consistency for Vision-Language Pre-training (CVPR 2024 · similarity 0.88)
- SineProject: Machine Unlearning for Stable Vision-Language Alignment (CVPR 2026 · similarity 0.87)
- Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment (ICLR 2025 · similarity 0.87)
- FFF: Fixing Flawed Foundations in Contrastive Pre-Training Results in Very Strong Vision-Language Models (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
