Research map / Vision-language and multimodal / Cross-modal learning
VALOR: Vision-Audio-Language Omni-Perception Pretraining Model and Dataset
Jing Liu, Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang
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 audio modality and VALOR-1M pretraining (VALOR_B vs vision-only VALOR_B^-) | VALOR_B · VALOR-32K retrieval | R@1: 43.3 -> 67.9 | Improved |
| add audio modality and VALOR-1M pretraining | VALOR_B · MSRVTT retrieval | R@1: 36.2 -> 43.0 | Improved |
| add audio modality and VALOR-1M pretraining | VALOR_B · VALOR-32K audiovisual captioning | CIDEr: 44.3 -> 55.7 | Improved |
| use dual softmax (DSL) post-processing at evaluation | VALOR_L · VALOR-32K retrieval | R@1: 73.2 -> 80.9 | Improved |
| pretrain on VALOR-1M instead of WebVid-2.5M (ablation) | VALOR (Video Swin-S) · MSRVTT retrieval | R@1: 35.7 -> 39.5 | Improved |
| pretrain on VALOR-1M instead of HD_VILA_10M (ablation) | VALOR (Video Swin-S) · MSRVTT retrieval | R@1: 30.4 -> 39.5 | Improved |
| add T-AV modality group to T-V only MGA pretraining (ablation) | MGA pretraining · VALOR-32K T-AV retrieval (finetune) | R@1: 46.9 -> 55.4 | Improved |
| use only T-AV group instead of T-V on T-V benchmark (ablation) | MGA pretraining · VALOR-32K T-V (finetune) | R@1: 42.7 -> 40.1 | Hurt |
| add T-V, T-A and T-AV groups together vs T-AV alone (ablation) | MGA pretraining · VALOR-32K T-A (finetune) | R@1: 12.3 -> 17.0 | Improved |
| add V-A, A-TV, V-TA groups on top of T-V/T-A/T-AV (ablation) | MGA pretraining · VALOR-32K T-AV (finetune) | R@1: 55.6 -> 54.0 | Hurt |
| use fine-grained plus weighted average instead of coarse score fusion (ablation) | MGA T-AV · AVR VALOR-32K | R@1: 53.4 -> 55.4 | Improved |
| use concatenated cross attention instead of merge attention (ablation) | MGC T-AV decoder · VALOR-32K AVC | CIDEr: 48.2 -> 50.0 | Improved |
Closest papers by meaning
- JointAVBench: A Benchmark for Joint Audio-Visual Reasoning Evaluation (ICLR 2026 · similarity 0.89)
- OmniVideoBench: Towards Audio-Visual Understanding Evaluation for Omni MLLMs (ICLR 2026 · similarity 0.88)
- MMIE: Massive Multimodal Interleaved Comprehension Benchmark for Large Vision-Language Models (ICLR 2025 · similarity 0.88)
- LoMo: Local Modality Substitution for Deeper Vision-Language Fusion (NeurIPS 2026 · similarity 0.88)
- OmniBench: Towards The Future of Universal Omni-Language Models (NeurIPS 2025 · similarity 0.88)
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · 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.
