Research map / Vision-language and multimodal / Cross-modal learning
OmniVinci: Enhancing Architecture and Data for Omni-Modal Understanding LLM
Hanrong Ye, Chao-Han Huck Yang, Arushi Goel, Wei Huang, Zhen Wan, Jinchuan Tian, An-Chieh Cheng, Ligeng Zhu, Yuanhang Su, Yuming Lou, Yong-Xiang Lin, Dong Yang and 20 more
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 Temporal Embedding Grouping to token concatenation (ablation) | OmniVinci (10B-token ablation) · Omni average | average accuracy: 45.51 -> 47.72 | Improved |
| use learned time embedding for absolute timestamps (ablation) | TEG baseline · Omni average | average accuracy: 47.72 -> 47.30 | No clear effect |
| use RoTE for absolute timestamps (ablation) | TEG baseline · Omni average | average accuracy: 47.72 -> 47.80 | No clear effect |
| add Constrained Rotary Time Embedding (ablation) | TEG baseline · Omni average | average accuracy: 47.72 -> 50.25 | Improved |
| add OmniAlignNet on top of TEG and CRTE (ablation) | OmniVinci (10B-token ablation) · Omni average | average accuracy: 50.25 -> 52.59 | Improved |
| add implicit learning from video audio (ablation) | visual-only model · VideoMME w/o subtitles | accuracy: 61.67 -> 63.76 | Improved |
| add implicit learning from video audio (ablation) | visual-only model · VideoMME Short | accuracy: 74.22 -> 71.31 | Hurt |
| add explicit learning with omni-modal data engine (ablation) | Visual + Audio (IL) · VideoMME w/o subtitles | accuracy: 63.76 -> 67.37 | Improved |
| add GRPO post-training with omni-modal MCQ data (ablation) | OmniVinci · Omni average | average accuracy: 53.73 -> 54.52 | Improved |
| include audio input in GRPO training (ablation) | OmniVinci RL · GRPO accuracy reward | accuracy reward: +0.1 higher convergence | Improved |
| apply AWQ 4-bit quantization | OmniVinci · Tennis MCQ | inference time (s): 3.29 -> 1.85 | Improved |
Later papers that compared with this method
1 later accepted paper reports OmniVinci as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- LatentOmni: Rethinking Omni-Modal Understanding via Unified Audio-Visual Latent Reasoning (NeurIPS 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- NVILA — NVILA: Efficient Frontier Visual Language Models (CVPR 2025) · better on 3 of 3 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
- OmniBench: Towards The Future of Universal Omni-Language Models (NeurIPS 2025 · similarity 0.91)
- OmnixR: Evaluating Omni-modality Language Models on Reasoning across Modalities (ICLR 2025 · similarity 0.90)
- Omni-Interactive Universal Embedder (NeurIPS 2026 · similarity 0.90)
- Omni-Captioner: Data Pipeline, Models, and Benchmark for Omni Detailed Perception (ICLR 2026 · similarity 0.89)
- LatentOmni: Rethinking Omni-Modal Understanding via Unified Audio-Visual Latent Reasoning (NeurIPS 2026 · similarity 0.89)
- A Benchmark for Omni-Modal Reasoning in Long Videos (NeurIPS 2026 · similarity 0.89)
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.
