Research map / Vision-language and multimodal / Multimodal LLMs
V2PE: Improving Multimodal Long-Context Capability of Vision-Language Models with Variable Visual Position Encoding
Junqi Ge, Ziyi Chen, Jintao Lin, Jinguo Zhu, Xihui Liu, Jifeng Dai, Xizhou Zhu
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 |
|---|---|---|---|
| fine-tune on augmented long-context data instead of short-context only (ablation) | InternVL2-2B · Long-VQA / MM-NIAH | accuracy | Improved |
| reduce positional increment of visual tokens to 1/16 or 1/256 (fixed) (ablation) | InternVL2-FT-32K · Long-VQA / MM-NIAH | accuracy | Improved |
| add V2PE (variable visual position increments) in training (ablation) | InternVL2-2B · MM-NIAH image retrieval, 256K tokens | accuracy: 1.5 -> 56.9 | Improved |
| add second-stage 256K fine-tuning with V2PE (ablation) | InternVL2-V2PE-32K · MM-NIAH image retrieval, 512K tokens | accuracy: 10.8 -> 95.1 | Improved |
| apply V2PE to short-context benchmarks with small increments (ablation) | InternVL2-V2PE · DocVQA, InfoVQA | accuracy | Hurt |
| apply V2PE to visual tokens only versus textual tokens or both (ablation) | InternVL2-V2PE-32K · MM-NIAH (six tasks) | accuracy | Improved |
| use V2PE instead of visual token compression (ablation) | InternVL2-V2PE-32K · MM-NIAH image retrieval | accuracy | Improved |
| use variable positional increments instead of fixed during training (ablation) | InternVL2-FT-32K · MM-NIAH image retrieval | accuracy | Improved |
| use V2PE instead of linear interpolation or NTK-aware RoPE extension (ablation) | InternVL2-FT-32K · Long-VQA / MM-NIAH | accuracy | Improved |
| use V2PE-32K versus InternVL2-2B | InternVL2-2B · MM-NIAH Avg | accuracy: 21.0 -> 70.2 | Improved |
| use V2PE-32K versus InternVL2-2B | InternVL2-2B · MileBench Avg | accuracy: 49.9 -> 55.9 | Improved |
| use V2PE-32K versus InternVL2-2B on short-context | InternVL2-2B · General benchmarks average | accuracy: 72.4 -> 73.0 | No clear effect |
Closest papers by meaning
- HoPE: Hybrid of Position Embedding for Long Context Vision-Language Models (NeurIPS 2025 · similarity 0.92)
- MMLongBench: Benchmarking Long-Context Vision-Language Models Effectively and Thoroughly (NeurIPS 2025 · similarity 0.91)
- Revisiting Multimodal Positional Encoding in Vision–Language Models (ICLR 2026 · similarity 0.91)
- MMLongCite: A Benchmark for Evaluating Faithfulness of Long-Context Vision-Language Models (NeurIPS 2026 · similarity 0.91)
- Towards General Continuous Memory for Vision-Language Models (NeurIPS 2025 · similarity 0.90)
- POINTS-Long: Adaptive Dual-Mode Visual Reasoning in MLLMs (CVPR 2026 · similarity 0.90)
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.
