Polish Tables & Figures

Turn draft-quality LaTeX floats into the tables and figures reviewers expect to see at ACM / IEEE / NeurIPS-style / LNCS venues: booktabs rules, columns that fit the page budget, clean subfigure layouts, venue-correct captions, consistent \cref references, and colorblind-safe colors. Every change is presentation-only; the numbers are sacred.

When to use

• "Make my tables/figures look professional" / "polish for camera-ready"
• "Convert this table to booktabs" / "my table is too wide" / "overfull hbox"
• "Lay these plots out as subfigures" / "fix my captions"
• "Are my figure colors colorblind-safe?" / "pick a palette for this plot"
• "Fix my \ref/\cref mess"
• After tailor-to-venue, before preflight-check.

Inputs

1. The main .tex file (the one with \documentclass); \input/\include files are followed automatically.
2. Optional: a venue profile venues/conferences/<venue>-<year>.yml (schema in venues/schema.yml) — sets column count and caption conventions. Without one, conventions are inferred from the documentclass.
3. Strongly recommended: a compiled .log next to the .tex so the column-sizing (overfull box) check can run.

Process

1. Resolve venue conventions. If a venue is named, load its profile (sanity-check with python3 scripts/venueyaml.py venues/conferences/<venue>.yml). Re-verify format-critical facts against the live cfp_url before relying on them: template/documentclass, column count, any stated figure/table rules (minimum font sizes, color/grayscale requirements, accessibility requirements). Profiles are a starting point, never ground truth.

2. Lint everything first:

   python3 scripts/check_floats.py paper.tex \
       --venue venues/conferences/<venue>-<year>.yml
   

   Flags: --json, --strict (warnings also fail), --log <file>, --overfull-threshold <pt>, --no-inputs. Exit codes: 0 clean, 1 findings, 2 bad arguments. The linter covers: booktabs style (\hline, \cline, vertical rules, double rules, missing package), \resizebox/tiny-font scaling, caption position per family, missing \caption/\label/\centering/ACM \Description, label-before-caption, deprecated subfigure package, width=\textwidth in a one-column float at a two-column venue, raster graphics, [h]/[H] placement, undefined and unreferenced labels, mixed Figure~\ref vs \cref styles, "Figure \cref" double prefixes, cleveref load order, lowercase \cref at sentence start, and Overfull \hbox entries from the .log.

3. Fix tables — work through findings using references/tables.md: the mechanical booktabs conversion recipe (it preserves cell content byte-for-byte), then column sizing within the page budget (cut → abbreviate → align numbers → \tabcolsep → tabularx → table* — never \resizebox).

4. Fix figures — references/figures.md: width discipline (\columnwidth/\linewidth, figure* for spans), subfigure layout with subcaption, vector formats, font sizes inside plots, float placement.

5. Fix captions and cross-references — references/captions-crossrefs.md: per-family caption position/style table, self-contained caption writing, and the cleveref setup (load last, capitalise, \cref vs \Cref, one style everywhere).

6. Fix color — references/color-accessibility.md. Check the colors actually used in the paper (hex values from TikZ/pgfplots definitions, or sampled from included plots):

   python3 scripts/palettes.py check 4477AA EE6677 ...   # arbitrary colors
   python3 scripts/palettes.py show okabe-ito --format latex|pgfplots|matplotlib
   

   check simulates protanopia/deuteranopia/tritanopia and grayscale print, flags confusable pairs (exit 1), and the references file explains redundant encoding (markers + line styles, not color alone). If plots are generated by user scripts (matplotlib/pgfplots), edit the plotting source and ask the user to regenerate — never claim recolored results.

7. Recompile and re-run the linter until the verdict is PASS (or every remaining WARN is a deliberate, user-approved choice). Then hand off to preflight-check for the full desk-reject lint.

Output

Edited .tex (and plotting-script) changes applied in place with a short change log per float: what changed and why, citing the venue convention. Plus the final linter report (PASS / PASS-WITH-WARNINGS / FAIL). No new files unless the user asks for extracted style snippets (e.g. a colors.tex palette preamble).

Adapt to your discipline

Conventions here target CS venues (ACM/IEEE/ML/LNCS). Forking for another field: adjust the caption-position conventions in references/captions-crossrefs.md and the family mapping in scripts/check_floats.py (convention_label); the booktabs, sizing, and color guidance is field-agnostic.

Guardrails

• Never change data: cell values, units, significant digits, axis scales, error bars, and legends' meanings are untouchable. Reformatting only. If a value looks wrong, tell the user — do not "fix" it.
• Never fabricate venue rules. The profile and the live CFP are the only sources; when neither states a rule, label advice as general convention.
• Do not regenerate or recolor result plots yourself; edit the user's plotting source and let them re-run it.
• Bold-best-result formatting must come from the user's stated criterion, not your reading of which number "looks best".
• Citation issues found along the way go to verify-citations; submission readiness goes to preflight-check. Never submit anything on the user's behalf.