Research map / Vision-language and multimodal / Multimodal LLMs
FALCON: Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers
Renshan Zhang, Rui Shao, Gongwei Chen, Miao Zhang, Kaiwen Zhou, Weili Guan, Liqiang Nie
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 ReAtten (ablation) | FALCON · V* | V*_Avg: 51.3 -> 61.3 | Improved |
| add ReAtten (ablation) | FALCON · MME-RealWorld Perception | Avg-C: 38.2 -> 42.1 | Improved |
| add ReAtten (ablation) | FALCON · MME-RealWorld Reasoning | Avg-C: 35.6 -> 39.0 | Improved |
| add ReAtten (ablation) | FALCON · POPE | accuracy: 85.7 -> 87.3 | Improved |
| use ReAtten instead of W-Atten shifted window attention (ablation) | FALCON · V* | V*_Avg: 60.2 -> 61.3 | Improved |
| use Complementary Image Pyramid (CIP) instead of no continuity method (ablation) | FALCON baseline · V* | V*_Avg: 51.3 -> 50.3 | Hurt |
| use CIP instead of no continuity method (ablation) | FALCON baseline · MME-RealWorld Reasoning | Avg-C: 35.6 -> 37.2 | Improved |
| train FALCON on LLaVA-1.5 data instead of LLaVA-v1.5-7B | FALCON · TextVQA | accuracy: 58.2 -> 61.3 | Improved |
| train FALCON on LLaVA-1.5 data instead of LLaVA-v1.5-7B | FALCON · MM-Vet | score: 31.1 -> 31.4 | No clear effect |
| increase visual registers from 36 to 64 to 144 (ablation) | FALCON · MME-RealWorld / V* (registers) | accuracy: 1.15 and 2.41 times more training time | Mixed |
| use ReCompact instead of pooling, Pixel Shuffle or Abstractor (ablation) | FALCON · MME-RealWorld Avg-C sum | Avg-C | Improved |
| Abstractor compression instead of pooling (ablation) | FALCON · MME-RealWorld Avg-C sum | Avg-C | Hurt |
Closest papers by meaning
- Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders (ICLR 2026 · similarity 0.90)
- Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · similarity 0.89)
- LION: Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge (CVPR 2024 · similarity 0.89)
- VIGC: Visual Instruction Generation and Correction (AAAI 2024 · similarity 0.88)
- Unleashing the Intrinsic Visual Representation Capability of Multimodal Large Language Models (CVPR 2026 · similarity 0.88)
- Visual-Redundancy-Controlled Parallel Decoding for Diffusion-Based Multimodal Large Language Models (NeurIPS 2026 · similarity 0.88)
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.
