arXiv new/cross; visual-token grounding for MLLM reasoning · P2 · 2026-07-09

Segmentation before Answering：它和通用视觉自监督的关系在于：把 MLLM 的 zoom-in 单位从 box 换成 mask，并讨论 segmentation patches 与 visual tokens 的对齐

中相关；详见方法、贡献和实验边界。

Segmentation before Answering: Pixel Grounding for MLLM Visual Reasoning　arXiv new/cross; visual-token grounding for MLLM reasoning　原文链接

编号2607.05798 优先级P2 类别arXiv new/cross; visual-token grounding for MLLM reasoning 会议arXiv new/cross; visual-token grounding for MLLM reasoning 方法把 MLLM 的 zoom-in 单位从 box 换成 mask，并讨论 segmentation patches 与 visual tokens 的对齐来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：把 MLLM 的 zoom-in 单位从 box 换成 mask，并讨论 segmentation patches 与 visual tokens 的对齐。中相关；详见方法、贡献和实验边界。

**Figureure 2 · Overview of our SegAnswer method**Figure 2. Overview of our SegAnswer method. The training pipeline progresses through three stages: Stage 1: Pixel Grounding aligns textual semantics with pixel-level features, training the MLLM to generate segmentation masks via a specialized <|seg|> token and a mask decoder. Stage 2: Multimodal Interleaved SFT enables the model to employ segmentation as an intermediate conversation step, using the generated mask to focus the visual context before answering. Stage 3: Reasoning with Pixel Grounding utilizes reinforcement learning to enhance MLLM visual reasoning by precise and finer segmented visual inputs.这张图概括 Segmentation before Answering 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。

**(b) Our segmentation-based pixel grounding**(b) Our segmentation-based pixel grounding. Figure 1. Comparison between the BBox-based zoom-in operation and our segmentation-based pixel grounding. (a) Rectangular bounding boxes inevitably introduce redundant background regions (e.g., the background around the sloping tennis racket) and fail to precisely disentangle the region of interest from overlapping objects (e.g., the fork and napkin), leading to visual noise and semantic ambiguity. (b) Pixel-level segmentation can precisely isolate the region of interest, effectively eliminating background noise and decoupling adjacent entities. In addition, by keeping the position index of the original image, sparse segmented image patches can also effectively reflect spatial relations.这张图/表用于判断 Segmentation before Answering 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致，而不是只看单个最高分。

核心问题

它和通用视觉自监督的关系在于：把 MLLM 的 zoom-in 单位从 box 换成 mask，并讨论 segmentation patches 与 visual tokens 的对齐。

方法拆解

把 MLLM 的 zoom-in 单位从 box 换成 mask，并讨论 segmentation patches 与 visual tokens 的对齐

主要贡献

中相关；详见方法、贡献和实验边界。

实验看点

实验部分建议重点看两类证据：一是作者是否把方法收益和更强数据、更长训练、更大模型区分开；二是跨模型、跨数据或跨任务迁移是否还能保留同样趋势。

局限与读法

这篇论文的结论需要结合任务设置、训练数据规模和消融实验一起看；不要只凭单个指标判断它对通用视觉表征的价值。