Flamingo: a Visual Language Model for Few-Shot LearningFlamingo: a Visual Language Model for Few-Shot Learning
问题:多模态模型如何用少量标注示例快速适配新任务,并处理任意交错的图文序列。
证据位置:摘要给出三条架构创新:桥接预训练视觉/语言模型、处理交错图文、摄入图像或视频;在大量图像与视频任务上 few-shot SOTA;声称在多个基准上超过用数千倍任务数据微调的模型。
作者自陈局限:模型与数据规模大且未开源;评测集中在图像/视频任务;few-shot 仍需任务示例;开放域安全与偏见未讨论。
它没主张的:不是零样本通用助手;不是小模型方案;不是无需任务示例。Problem: how can a multimodal model adapt to new tasks from a handful of examples and handle arbitrarily interleaved image-text sequences?
Where the evidence sits: the abstract lists three architectural innovations — bridging pretrained vision-only and language-only models, handling interleaved image-text, and ingesting images or videos; it reports few-shot SOTA across image and video tasks and claims to outperform models fine-tuned on thousands of times more task data on several benchmarks.
What the authors say about limits: large model and data scale, not open-sourced; evaluation centres on image/video tasks; few-shot still needs task examples; open-domain safety and bias are not discussed.
What it does not claim: not a zero-shot general assistant; not a small-model recipe; not free of task examples.
关键参数Key specs
- 证据位置 / evidence
- 摘要:三条架构创新与 few-shot SOTA
- 模态 / modalities
- 图像、视频、交错图文
- 复现状态 / reproduction
- 未复现(只读摘要)
官方原文怎么说What the source says
bridge powerful pretrained vision-only and language-only models
handle sequences of arbitrarily interleaved visual and textual data
a single Flamingo model can achieve a new state of the art with few-shot learning
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-19
- 许可 / 类型Licence / kind
- arxiv-preprint
- 许可说明Licence evidence
- https://arxiv.org/abs/2204.14198
- 类目Category
- 多模态Multimodal
论文里的数字怎么读How to read the numbers
先分清测什么、怎么算、和谁比、有没有数据污染。拿不准就回原文看实验设置。Before citing a score, separate what it measures, how it is scored, what it compares against and whether contamination is controlled. If unsure, go back to the experiment setup in the paper. 看评测基准See benchmark cards · 看数据集卡See dataset cards