[于 2026 年 9 月 7 日提交]

标题：面向科学图表理解的结构标记证据锚定推理 (Structure-Token Evidence-Anchored Reasoning for Scientific Chart Understanding)

查看由 Alberlucia Rafael Soarez 及其他 4 位作者撰写的论文《Structure-Token Evidence-Anchored Reasoning for Scientific Chart Understanding》PDF 版本

查看 PDF HTML (实验性)

摘要：科学图表通过坐标轴、图例和几何标记来编码数量信息，但现有大型视觉语言模型（LVLM）仍将其视为普通自然照片。视觉上下文示例无法体现坐标系框架；无约束的思维链（Chain-of-Thought）常常凭空给出一个看似合理、但从未在柱状图中实际读取过的数值。我们提出了 STEER（Structure-Token Evidence-anchored Reasoning，结构标记证据锚定推理），该方法冻结了 Llama-3.2-Vision 编码器并引入三个模块：用于绑定刻度、图例项与几何标记的图表结构图编码器（CSGE）；强制每一个算术计算步骤都必须引用图节点的证据锚定单步推理（EASR）；以及仅将专用表格提取器用作节点属性教师的弱解析器-强推理器对齐机制（WPSR）。在 ChartQA 基准上，STEER 达到了 82.70 的平均松弛准确率（relaxed accuracy），高于 ChartGemma 的 80.16 以及在相同混合数据上训练的 LLaVA-CoT 骨干网络的 76.40。在无法依赖 OCR 捷径的基准上，优势进一步扩大：CharXiv 推理评测达到 33.60（对比 InternVL Chat V1.5 的 29.20），ChartQAPro CoT 达到 40.70（对比 Qwen2-VL-7B 的 37.17）。消融实验表明，去除节点序列化或数值候选约束会导致绝大部分推理增益消失。

提交历史

来自：Alberlúcia Rafael Soarez [

查看邮箱

]

[v1] 2026 年 9 月 7 日，星期一 16:49:15 UTC (1,174 KB)

[Submitted on 7 Sep 2026]

Title:Structure-Token Evidence-Anchored Reasoning for Scientific Chart Understanding

View a PDF of the paper titled Structure-Token Evidence-Anchored Reasoning for Scientific Chart Understanding, by Alberlucia Rafael Soarez and 4 other authors

View PDF HTML (experimental)

Abstract:Scientific charts encode quantities in axes, legends, and geometric marks, yet large vision-language models still treat them as natural photographs. Visual in-context examples do not expose the coordinate frame; unconstrained chain-of-thought can name a plausible number that was never read from a bar. We present STEER (Structure-Token Evidence-anchored Reasoning), which freezes a Llama-3.2-Vision encoder and inserts three modules: a chart structure graph encoder (CSGE) that binds ticks, legend items, and marks; evidence-anchored step reasoning (EASR) that forces every arithmetic step to cite a graph node; and weak-parser strong-reasoner alignment (WPSR) that uses a specialized table extractor only as a teacher of node attributes. On ChartQA, STEER reaches 82.70 average relaxed accuracy versus 80.16 for ChartGemma and 76.40 for a LLaVA-CoT backbone trained on the same mix. Gains widen on CharXiv reasoning (33.60 vs. 29.20 InternVL Chat V1.5) and ChartQAPro CoT (40.70 vs. 37.17 Qwen2-VL-7B), where OCR shortcuts disappear. Ablations show that dropping node serialization or numeric candidate constraints undoes most of the reasoning lift.

Submission history

From: Alberlúcia Rafael Soarez [

view email

]

[v1]

Mon, 7 Sep 2026 16:49:15 UTC (1,174 KB)