智能降尺度：探究小型多模态模型中的感知与推理瓶颈
Downscaling Intelligence: Exploring Perception and Reasoning Bottlenecks in Small Multimodal Models
论文把多模态小模型的性能下降拆解为感知瓶颈与推理瓶颈，并提出 visual extraction tuning：先显式训练模型抽取与指令相关的视觉细节，再做分步推理。做 VLM 压缩、蒸馏或小模型部署的研究者可直接复用这一「先抽取视觉细节、再推理」的两阶段范式，并借用其感知/推理解耦的评测思路定位自身模型的失效环节。
原文依据 we introduce visual extraction tuning, which explicitly trains the model to extract instruction-relevant visual details consistently across tasks
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论扩大模型规模使多模态模型在视觉理解与推理方面取得了显著进展，但实际需求呼唤更小、更高效的系统。在本工作中，我们对多模态模型中的智能降尺度进行原理性分析，考察大语言模型（LLM）容量下降如何影响多模态能力。我们的初步发现揭示了一个有趣的趋势：LLM 的降尺度对视觉能力的影响不成比例地大，而对从 LLM 继承而来的能力影响较小。随后我们考察这一下降主要反映的是视觉推理能力如预期般的衰退，还是感知能力的更根本性丧失。在将 LLM 降尺度对感知的影响单独分离出来后，我们发现性能仍然急剧下降，其影响往往与对推理的影响相当甚至更大。为应对这一瓶颈，我们提出视觉抽取调优，显式训练模型跨任务一致地抽取与指令相关的视觉细节。借助这些抽取出的视觉细节，我们随后应用分步推理来生成答案。这些组件共同构成了我们的 Extract+Think 方法，为该领域树立了效率与性能的新标准。
英文原文摘要 1197 字符
Scaling up multimodal models has enabled remarkable advances in visual understanding and reasoning, but practical demands call for smaller, efficient systems. In this work, we conduct a principled analysis of downscaling intelligence in multimodal models, examining how reduced large language model (LLM) capacity affects multimodal capabilities. Our initial findings reveal an interesting trend: LLM downscaling disproportionately affects visual capabilities, rather than abilities inherited from the LLM. We then examine whether this drop mainly reflects the expected decline in visual reasoning or a more fundamental loss of perceptual abilities. Isolating the effect of LLM downscaling on perception, we find performance still drops sharply, often matching or exceeding the impact on reasoning. To address this bottleneck, we introduce visual extraction tuning, which explicitly trains the model to extract instruction-relevant visual details consistently across tasks. With these extracted visual details, we then apply step-by-step reasoning to generate answers. Together, these components form our Extract+Think approach, setting a new standard for efficiency and performance in this space.