01 · CLOSEST MATCHarXiv:2606.10334 · 2026-06-09

执行 student 代码,再用 render 教回 code

Visual-SDPO 把 screenshot、rubric 和 error 当 teacher-only context。

78.6

ChartMimic

82.6

Design2Code

29%

GRPO render budget

10/10

个人兴趣

来源:Abstract;Tables 1-31 / 9
02 · FAILUREarXiv:2606.10334 · 2026-06-09

Executable code 仍会生成明显错误的 artifact

Overlap

元素重叠

Clipping

文字被裁

Alignment

布局错位

Contrast

可读性不足

来源:Introduction2 / 9
03 · LOOParXiv:2606.10334 · 2026-06-09

训练环闭合在 student 自己的 artifact 上

1Code

student on-policy

2Execute

renderer/browser

3Observe

screenshot/error

4Teacher

score same code

5Update

KL + GRPO

来源:Method overview3 / 9
04 · CREDITarXiv:2606.10334 · 2026-06-09

不是每个 code token 都该为 defect 负责

w(token∈stmt) = 1 + (α−1) · max IoU(stmt regions, defects)

visual region 先映射到 statement,再放大 implicated token 的 distillation weight。

来源:Visual-Grounded Code Credit Weighting4 / 9
05 · INSTRUMENTarXiv:2606.10334 · 2026-06-09

三类 renderer 使用不同 source map

Matplotlib

artist hook + stack frame + bbox

HTML

source metadata + browser rectangle

Slides

python-pptx shape construction hook

来源:Method implementation5 / 9
06 · RESULTarXiv:2606.10334 · 2026-06-09

Visual feedback 稳定超过 reference-code OPSD

BenchmarkOPSDGRPOVisual-SDPO
ChartMimic77.076.278.6
Design2Code78.680.082.6
AeSlides—58.260.7
根据 Tables 1-3 重绘6 / 9
07 · FAILED EXECUTIONarXiv:2606.10334 · 2026-06-09

代码跑不起来,也还有学习信号

GRPO reward

execution failure → 0

Teacher context

traceback / browser error 仍可输入

来源:Objective description7 / 9
08 · GAME TRANSLATIONarXiv:2606.10334 · 2026-06-09

把 region→statement 扩展成 event→code span

Compile

diagnostic → syntax/type span

Gameplay

event/test → signal/state code

Visual

frame diff → node/property

本文实验设计建议8 / 9
09 · REPRODUCTIONarXiv:2606.10334 · 2026-06-09

最值得做,也最需要补齐公开性

论文给出

principle、formula、benchmark gains

论文缺少

code、α/β、optimizer、detector details

兴趣程度 10/109 / 9