tasks/DOC/DOC_task_0_latex_layout_fix.md

metadata

id: DOC_task_0_latex_layout_fix
name: TeXstudio 修复 LaTeX 排版溢出
category: DOC
timeout_seconds: 1500

Prompt

⚙️ Execution convention: this task is graded purely on output files — nobody is in the loop to approve anything. Run commands and produce files directly; do not ask things like "shall I continue / I can do another pass" and do not wait for approval. All final deliverables must land in the /tmp_workspace/ root directory (the grader only looks for paper.tex / paper.pdf / pdf_pN.png / changes.md / log_before.png directly under that directory — subdirectories don't count). When done, run ls -la /tmp_workspace/ to self-check before exiting.

Background: /tmp_workspace/paper.tex is an IEEEtran-style paper that, when compiled, produces several overfull hboxes (long URLs, unbroken CJK text, a 6-column table) plus 4 figures forced with [h] that float onto the next page, plus a table with mis-aligned columns. You need to fix it so it compiles cleanly with reasonable layout.

Task goal: edit paper.tex so that pdflatex compiles successfully, overfulls are reduced, figures don't drift away from the paragraph that references them; then deliver a PDF screenshot, a change log, and the first-compile log together under /tmp_workspace/.

Hard constraints on paper.tex

• pdflatex -interaction=nonstopmode paper.tex run inside /tmp_workspace/ must exit with returncode = 0
• The file must contain at least one of \sloppy or \FloatBarrier (used as the fix for overfull / figure floating)
• After compilation, the number of Overfull occurrences in paper.log must be < 3
• The compiled paper.pdf must have ≥ 9 pages

Companion deliverables (all in the /tmp_workspace/ root)

File	Requirement
pdf_p3.png / pdf_p6.png / pdf_p9.png	Screenshots of pages 3 / 6 / 9 of the fixed PDF. Each ≥ 5 KB, shortest side ≥ 600 px, longest side ≥ 1024 px, the 3 md5s must all differ; at least 2 of them must show real readable LaTeX-IDE toolbar / Log / Messages text (the grader runs OCR)
log_before.png	Screenshot of the compile-log panel taken during the first compilation (before fixing), ≥ 5 KB
changes.md	Change log, at least 20 list items each starting with \n- , each explaining what was changed and why

Suggested fix directions (not mandatory, pick as needed)

• Long URL overflow → \url{} + local \sloppy
• Unbroken CJK text → \hspace{0pt} / \,
• Mis-aligned table columns → switch to tabularx / \arraybackslash
• Figures with forced [h] floating off page → change to [!htbp] + \FloatBarrier

Expected Behavior

设计意图与典型解题路径（仅供出题人参考，不发给 agent）：

1. 推荐打开 TeXstudio 加载 paper.tex（也可以用 VS Code + LaTeX Workshop / Overleaf 风格命令行 + 手工编辑器，只要满足产物硬约束即可）。
2. 第一次编译并截 Log 面板 → log_before.png，统计 overfull / underfull 数量做基线。
3. 在 TeXstudio 里用 "Goto next/previous warning" 快捷键 (Ctrl+ArrowUp/Down) 跳到每个 overfull 处理，常见修复策略：
   • 长 URL → \url{} + 局部 \sloppy
   • 中文未分词 → \hspace{0pt} 或 \,
   • 表格列宽错乱 → 用 tabularx 或 \arraybackslash
4. 把 figure 的 [h] 改成 [!htbp] + \FloatBarrier，避免浮动跑出 section。
5. 表格可用 TeXstudio 的 "Wizards → Table" 重排，目标列对齐 lcr。
6. 反复编译直到 paper.log 中 Overfull < 3 且 figure 不再漂离引用段落。
7. 截 PDF 第 3 / 6 / 9 页 → pdf_p3.png / pdf_p6.png / pdf_p9.png。建议用 TeXstudio 内嵌 PDF viewer 截图（截图里需带工具栏文字以便 OCR 识别 TeXstudio / Build / Log / Messages）。
8. 写 changes.md：每条以 \n- 开头，列出 ≥ 20 项修改 + 行号 + 修复策略。

评分要点（hard gates）：

• \sloppy 或 \FloatBarrier 没出现 → 总分 cap 0.40
• pdflatex 编译失败 → cap 0.40
• PDF 不足 9 页 → cap 0.45
• 截图缺张 / 尺寸不达标 / md5 不唯一 → cap 0.45
• OCR 命中 TeXstudio/Build/Log/Messages 的截图数 < 2 → cap 0.50
• VLM 不可用 → cap 0.60
• VLM 排版评分 < 0.4 → cap 0.30；< 0.6 → cap 0.45

Source

• Reddit: https://www.reddit.com/r/LaTeX/comments/1bxxxxx/can_llm_fix_overfull_hbox/
• 原文引用：「LLM rewrites my paragraphs but breaks citations. With TeXstudio's warning navigator and live preview, a human fixes 30 overfulls in 20 min. Without GUI, agent gives up.」
• 对应 benchmark case: 学术排版 / DOC 新增

Grading Criteria

• paper.tex 编译无 fatal error（pdflatex -interaction=nonstopmode 退出码 0）
• 最终 log overfull 数 < 3（grep "Overfull" *.log）
• pdf_p3/6/9.png 三张存在，每张 ≥ 5KB、分辨率 ≥ 1024×600，且三张 md5 互不相同
• PDF 至少 9 页
• changes.md 列出 ≥ 20 处修改
• log_before.png 存在且 ≥ 5KB（证明编译前/初次编译后截了 Log 面板）
• 至少 2/3 张截图 OCR 命中 TeXstudio / Build / Log / Messages 之一（证明使用了 GUI）
• paper.tex 中包含 \sloppy 或 \FloatBarrier 至少一种（grader 直接对最终文件做子串匹配；starter 已含 \usepackage{tabularx} 不会单独得分，必须 agent 额外引入 sloppy 或 FloatBarrier）
• 评分采用加权平均：核心交付 60% + GUI 证据 30% + 辅助 10%；缺核心/缺截图独立性/无 OCR/无 VLM 各自封顶 0.40~0.60

Automated Checks

import subprocess, re, hashlib
from pathlib import Path
try:
    import pytesseract
except ImportError:
    pytesseract = None
from PIL import Image

def grade(workspace_path=None, **kwargs):
    workspace = Path(workspace_path) if workspace_path else Path("/tmp_workspace")
    r = {"checks": {}, "overall_score": 0.0}
    # bucketed sub-scores: core deliverable (0..1), gui evidence (0..1), aux (0..1)
    core_pts, core_max = 0, 4   # compile_ok, overfull<3, pages>=9, strategy_used
    gui_pts,  gui_max  = 0, 4   # screens exist+sized, screens md5 unique, tex_toolbar OCR, log_before
    aux_pts,  aux_max  = 0, 1   # changes>=20

    tex = workspace / "paper.tex"
    overfull_n = None
    if tex.exists():
        c = tex.read_text(errors="ignore")
        # starter already ships \usepackage{tabularx}? No — starter ships without it
        # in v3. Require agent to add \sloppy or \FloatBarrier explicitly.
        if any(k in c for k in ["\\sloppy", "FloatBarrier"]):
            r["checks"]["strategy_used"] = True; core_pts += 1
        try:
            p = subprocess.run(["pdflatex", "-interaction=nonstopmode", "paper.tex"],
                               cwd=str(workspace), capture_output=True, timeout=120)
            if p.returncode == 0:
                r["checks"]["compile_ok"] = True; core_pts += 1
            log = workspace / "paper.log"
            if log.exists():
                overfull_n = len(re.findall(r"Overfull", log.read_text(errors="ignore")))
                if overfull_n < 3:
                    r["checks"][f"overfull={overfull_n}<3"] = True; core_pts += 1
        except (FileNotFoundError, subprocess.TimeoutExpired, OSError):
            pass
        pdf = workspace / "paper.pdf"
        npages = 0
        try:
            from pdf2image import convert_from_path
            pgs = convert_from_path(str(pdf), dpi=80)
            npages = len(pgs)
        except Exception:
            try:
                info = subprocess.run(["pdfinfo", str(pdf)], capture_output=True,
                                      text=True, timeout=15).stdout
                for line in info.splitlines():
                    if line.startswith("Pages:"):
                        npages = int(line.split()[1]); break
            except Exception:
                pass
        if npages >= 9:
            r["checks"]["pages>=9"] = True; core_pts += 1

    # --- GUI evidence: screenshots must exist, be sized, md5-unique ---
    shots = [workspace / f"pdf_p{n}.png" for n in (3, 6, 9)]
    if all(p.exists() for p in shots):
        sized = True
        for p in shots:
            try:
                if p.stat().st_size < 5 * 1024:
                    sized = False; break
                with Image.open(p) as im:
                    w, h = im.size
                if w < 1024 or h < 600:
                    sized = False; break
            except Exception:
                sized = False; break
        if sized:
            r["checks"]["screens_sized"] = True; gui_pts += 1
        try:
            md5s = {hashlib.md5(p.read_bytes()).hexdigest() for p in shots}
            if len(md5s) == 3:
                r["checks"]["screens_unique"] = True; gui_pts += 1
        except Exception:
            pass

    cm = workspace / "changes.md"
    if cm.exists() and cm.read_text(errors="ignore").count("\n- ") >= 20:
        r["checks"]["changes>=20"] = True; aux_pts += 1

    # log_before.png: evidence that agent compiled once and screenshotted Log panel
    log_before = workspace / "log_before.png"
    if log_before.exists():
        try:
            if log_before.stat().st_size >= 5 * 1024:
                r["checks"]["log_before"] = True; gui_pts += 1
        except Exception:
            pass

    # OCR for TeXstudio toolbar: require ≥2 of the 3 screenshots to hit a
    # TeXstudio keyword (raises the bar vs. the v1 single-screenshot check).
    if pytesseract is not None:
        hits = 0
        for p in shots:
            if not p.exists():
                continue
            try:
                tx = pytesseract.image_to_string(Image.open(p))
                if any(k in tx for k in ["TeXstudio", "Build", "Log", "Messages"]):
                    hits += 1
            except Exception:
                pass
        if hits >= 2:
            r["checks"]["tex_toolbar"] = True; gui_pts += 1
        r["checks"]["tex_toolbar_hits"] = hits

    core = core_pts / core_max
    gui  = gui_pts  / gui_max
    aux  = aux_pts  / aux_max
    base = round(0.6 * core + 0.3 * gui + 0.1 * aux, 3)

    # --- VLM rubric ---
    try:
        from _judge_helper import vlm_score_rubric
    except Exception:
        vlm_score_rubric = None
    pdf_imgs = [str(p) for p in shots if p.exists()]
    vlm_avg = None
    if vlm_score_rubric and pdf_imgs:
        rubric = {
            "vlm_no_overfull_text": "页面内容未出现明显超出页边的文字、表格或公式（无 overfull box）",
            "vlm_figure_aligned": "图表整齐位于其引用段落附近，未漂离上下文（不在错误的页码末尾）",
            "vlm_table_fits_page": "表格不溢出页面、列宽合理、表头清晰",
            "vlm_typography_normal": "正文段落对齐自然，无超长间距、不规则换行或单字行尾",
        }
        vlm = vlm_score_rubric(pdf_imgs[:3], rubric, instruction="评估 LaTeX 修复后的 PDF 排版质量。")
        for k in rubric:
            r["checks"][k] = vlm.get(k, 0.0)
        r["judge_method"] = vlm.get("judge_method", "failed")
        vlm_avg = sum(vlm.get(k, 0.0) for k in rubric) / len(rubric)
        zero_overfull = (overfull_n is not None and overfull_n == 0) and r["checks"].get("compile_ok", False)
        if zero_overfull:
            overall = round(0.7 * base + 0.3 * vlm_avg, 3)
        else:
            overall = round(0.5 * base + 0.5 * vlm_avg, 3)
    else:
        overall = base

    # --- Hard gates (multi-layer, raised vs. v1) ---
    # 1. Core deliverable broken → cap 0.40
    if not (r["checks"].get("compile_ok") and r["checks"].get("strategy_used")):
        overall = min(overall, 0.40)
    # 2. PDF too short → cap 0.45
    if not r["checks"].get("pages>=9"):
        overall = min(overall, 0.45)
    # 3. GUI evidence missing (no sized screenshots OR not md5-unique) → cap 0.45
    if not (r["checks"].get("screens_sized") and r["checks"].get("screens_unique")):
        overall = min(overall, 0.45)
    # 4. TeXstudio OCR not detected (only enforce when OCR backend available) → cap 0.50
    if pytesseract is not None and not r["checks"].get("tex_toolbar"):
        overall = min(overall, 0.50)
    # 5. VLM unavailable → cap 0.60 (cannot reach top tier without visual judge)
    if vlm_avg is None:
        overall = min(overall, 0.60)
    # 6. Strict VLM degradation
    if vlm_avg is not None:
        if vlm_avg < 0.4:
            overall = min(overall, 0.30)
        elif vlm_avg < 0.6:
            overall = min(overall, 0.45)

    r["overall_score"] = round(overall, 3)
    r["sub_scores"] = {"core": round(core, 3), "gui": round(gui, 3), "aux": round(aux, 3),
                       "base": base, "vlm": vlm_avg}
    return r

Workspace Path

workspace/DOC/task_0_latex_layout_fix/

Skills

Env

Warmup

which pdflatex >/dev/null 2>&1 || DEBIAN_FRONTEND=noninteractive apt-get install -y -qq texlive-latex-base texlive-latex-extra texlive-latex-recommended texlive-publishers texlive-fonts-recommended texlive-lang-chinese || true
which pdfinfo >/dev/null 2>&1 || DEBIAN_FRONTEND=noninteractive apt-get install -y -qq poppler-utils || true
which tesseract >/dev/null 2>&1 || DEBIAN_FRONTEND=noninteractive apt-get install -y -qq tesseract-ocr || true
pip install -q pdf2image pytesseract pillow numpy || true