Research map / Vision-language and multimodal / Multimodal LLMs
UniCode : Learning a Unified Codebook for Multimodal Large Language Models
Sipeng Zheng, Bohan Zhou, Yicheng Feng, Ye Wang, Zongqing Lu
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 |
|---|---|---|---|
| use a unified codebook instead of separate visual and text tokenizers (ablation) | UniCode · ImageNet generation | FID: 9.82 -> 6.72 | Improved |
| use a unified codebook instead of visual encoder plus text tokenizer (ablation) | UniCode · VQAv2 | accuracy: 52.3 -> 53.1 | Improved |
| use language-driven iterative codebook training instead of frozen LLM codebook (ablation) | UniCode · ImageNet generation | FID: 34.45 -> 6.72 | Improved |
| use dual alternative training instead of language-driven iterative training (ablation) | UniCode · VQAv2 | accuracy: 53.1 -> 9.3 | Hurt |
| add in-context image decompression pre-training task (ablation) | UniCode · ImageNet generation | FID: 7.08 -> 6.72 | Improved |
| upgrade visual tokenizer from VQ-GAN to RQ-VAE to HQ-VAE (ablation) | UniCode · POPE | accuracy: 65.1 -> 71.8 | Improved |
| add extra CC3M images, ground-truth images, and a pretrained larger ViT to the visual encoder (ablation) | UniCode · POPE | accuracy: 71.8 -> 77.6 | Improved |
| test at code map resolution larger than 256 used in pretraining (ablation) | UniCode · VQAv2 | accuracy: 53.1 -> 41.3 (320) | Hurt |
| pretrain at 320x320 to match test resolution (ablation) | UniCode · VQAv2 | accuracy: 41.3 -> 54.5 | Improved |
| use UniCode instead of HQ-TVAE* of same 7B scale | image generation · CC3M val | FID: 12.13 -> 11.54 | Improved |
| use unified codebook versus HQ-VAE for reconstruction | HQ-UniCode · ImageNet reconstruction | rFID: 2.61 -> 2.83 | No clear effect |
| compare UniCode against LLaVA-1.5 on VQA | UniCode · POPE | accuracy: 86.4 -> 71.8 | Hurt |
Later papers that compared with this method
1 later accepted paper reports Unicode as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Asymmetric Hierarchical Anchoring for Robust Audio–Visual Cross-Modal Generalization (NeurIPS 2026 · better on 9 of 9 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
- From Pixels to Tokens: Byte-Pair Encoding on Quantized Visual Modalities (ICLR 2025 · similarity 0.90)
- Vision as a Dialect: Unifying Visual Understanding and Generation via Text-Aligned Representations (NeurIPS 2025 · similarity 0.89)
- Mono-InternVL: Pushing the Boundaries of Monolithic Multimodal Large Language Models with Endogenous Visual Pre-training (CVPR 2025 · similarity 0.88)
- Unleashing the Intrinsic Visual Representation Capability of Multimodal Large Language Models (CVPR 2026 · similarity 0.88)
- Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inference (AAAI 2025 · similarity 0.88)
- Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · 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.
