CommunityRedacción y edicióngithub.com

humanize-ml-paper

Writing skill. Write or revise a full ML conference paper (ICLR, NeurIPS, ICML) in one consistent house style for ML conference papers, distilled from a set of published papers. Use when turning a research proposal, method notes or experiment results into a complete LaTeX paper; when drafting or rewriting any section (abstract, introduction, preliminary, methodology, experiments, limitations, conclusion, appendix, related works) in this house style; when planning page budgets or main-text vs appendix allocatio

¿Qué es humanize-ml-paper?

humanize-ml-paper is a Antigravity agent skill that writing skill. Write or revise a full ML conference paper (ICLR, NeurIPS, ICML) in one consistent house style for ML conference papers, distilled from a set of published papers. Use when turning a research proposal, method notes or experiment results into a complete LaTeX paper; when drafting or rewriting any section (abstract, introduction, preliminary, methodology, experiments, limitations, conclusion, appendix, related works) in this house style; when planning page budgets or main-text vs appendix allocatio.

Compatible con~Claude Code~Codex CLI~Cursor✓Antigravity
npx skills add https://github.com/LehengTHU/humanize-ml-paper/tree/HEAD/skills/humanize-ml-paper

Installed? Explore more Redacción y edición skills: steipete/notion, langchain-ai/langchain, bytedance/podcast-generation · View all 6 →

Preguntar en tu IA favorita

Abre un nuevo chat con esta habilidad de agente ya precargada.

Documentación

ML conference paper style

This skill reproduces one consistent house style for ML conference papers: the section skeleton, paragraph templates, transitions, notation, experiment write-up, figure palette, table layout and LaTeX project structure. The rules were distilled from a set of published ICLR / NeurIPS / ICML papers, and every rule carries an evidence label that says how strongly the source set supports it. All examples in this skill are synthetic.

TypeContribution (references/paper_types.md)
Wworkflow with components: new modules inside a training, agent or inference loop, trained end to end
Tobjective + constraint, solved in closed form; inference-time or training-free
Lnew training loss with a theoretical guarantee
Rreward or supervision design for RL
Ffindings: an empirical discovery, possibly followed by a simple method

Evidence labels. [all] = every source paper where the rule applies follows it; [most] = a clear majority; [some] = a minority or a single precedent: an option, never a checklist requirement on its own; [venue: ICML] / [venue: ICLR] / [venue: NeurIPS] = a venue requirement or practice seen only at that venue (all two-column guidance is [venue: ICML] and rests on the official ICML kit plus the house two-column layout); [current] / [older] = practice that changed over time (follow [current]); [fix] = the skill corrects what the sources did; [skill choice] = a template helper or default the skill adds. Unlabelled rules are [all] or [most].

When practice disagrees: an [all] or [most] convention is the style. Where practice changed over time, follow [current]. A [current] rule that another current source contradicts is a preference, not a checklist item. Venue requirements beat style, and the current call for papers beats everything in this skill (references/venues.md). Where the sources are wrong (a bug, a font fallback), the skill says it corrects them ([fix]).

Reference files: read them when you reach the step

FileRead when
references/venues.mdStep 0 (confirm the venue, fetch the current call); switching venues; camera-ready. It has limits, statements, citation commands, the ICML 2026 rules and tectonic build recipe, a one-column ↔ two-column conversion checklist, and what check_paper.py checks
references/paper_types.mdStep 0-1: pick the paper type and take its intro P2-P4 wording and Methodology plan (each type is a worked pattern [some])
references/structure.mdStep 1-2 and 5: arc, section titles, page/word blueprints (ICLR, NeurIPS, ICML), main vs appendix split, appendix order, statement templates
references/latex_source.mdStep 0 and whenever you create a file, label, figure file, macro or bib entry; before the camera-ready switch (revision macros)
references/concepts.mdStep 1 (build the concept table and the term table: borrow terms from the cited literature before coining, with a citation key and a quote for every key term, §3) and Step 3 (mark concepts, keep one phrase per concept); budget 2-5 core + ≤6 framing, 8-15 marks, Intro + first definition only; one orange \concept{}; core concepts echoed later
references/introduction.mdStep 3 before drafting the Introduction or Related Works: the argument chain to write first (§0), the 19-move skeleton with evidence from the house papers (§1), method vs findings papers (§2), how prior work is introduced and how RW paragraphs end (§4), draft → final revision patterns (§5), rules IR1-IR15 with checks (§6), a worked diagnosis of a method draft (§7)
references/contributions.mdStep 1 (pick the epithet) and Step 3-7 wherever a contribution is stated (abstract, teaser caption, method opener, RQ labels, limitations, conclusion, RW closers): devices D1-D18, verb tiers by evidence, conclusion / limitation / number / outperform rules, flaws to avoid, draft → final revisions, how to state contributions while results are pending (C.7), rules KR1-KR14 with checks
references/prose.mdStep 3 before writing ANY paragraph: abstract and intro slot templates, transitions, connective library by function (§4a, four tiers, with real counts), cohesion rules (§4b: every sentence links back, density targets, double lock, RQ bridges, hand-offs), phrase bank, banned and older-style words, Limitations/Conclusion, Related Works arc, revision rules, tense and density norms, sentence shape / subject consistency / emphasis (§12, SR1-SR15), the 27-habit "avoid → use instead" table (with counts), and how to use the shipped lexicon (or build your own)
assets/lexicon/Step 3, before writing a sentence: lexicon.md / lexicon.json are a 1,272-entry phrase bank by writing step (M1 background ... M10 conclusion, connectives, word classes), each with its tier (core/occasional/retired/ref_only/banned) and an example; ships with the skill, no setup needed
tools/Step 3, to check any multi-word phrase: ps.py queries the shipped tools/phrase_index.json.gz by default (no setup); with your own papers, build_corpus.py builds a sentence corpus from YOUR LaTeX sources into your workspace and ps.py --corpus / ngrams.py search it instead (tools/README.md)
references/math.mdWriting the Preliminary/Methodology, any definition, lemma, theorem or proof, an algorithm block, or two-column displays
references/experiments.mdWriting the Experiments section, implementation details, analysis paragraphs, or an experiment plan with empty tables
references/figures_tables.mdMaking any plot, diagram, caption or table; colours, main-table variant, float placement; two-column float rules with the two-column float budget (§6.1); measured print-size diagram details: strokes, shadow, arrowheads, radii, padding, text, legend (§5.2)
references/examples.mdBefore drafting a section, to see a short annotated synthetic example of the target technique
assets/Step 3-4: lexicon/ is the phrase bank (see the references/prose.md row above); diagrams/, plots/ are synthetic reference diagrams and plots that show the conventions to imitate (listed in figures_tables.md §5); diagram_src/ holds the TikZ source of each diagram and style_sheet.png, every diagram style with its name
templates/Step 0: copy into the paper folder (see "Templates")

Workflow: proposal to finished paper

Step 0. Confirm the venue, fetch the current call for papers, set up

  1. Confirm the target venue and cycle with the user, then fetch that year's official call for papers, author instructions and style kit (venues.md §1.0). Limits and required statements change every year: ICLR went from 10 to 9 pp (2026) and added a required AI use statement (2027); NeurIPS 2026 added track options; ICML 2027 had no call yet on 2026-09-25. Write the limits and required statements as % comments at the top of the main file, and pass --limit to check_paper.py if they differ from its VENUES table. Never reuse last year's .sty/.bst.
  2. Collect the inputs: proposal or method notes, available results (maybe none), prior-work list.
  3. Decide the paper type (paper_types.md): W workflow, T objective + closed form, L loss + theory, R reward design, F findings. A proposal that introduces a method is a method paper.
  4. Create the project. ICLR/NeurIPS: copy templates/main_skeleton.tex (for NeurIPS, swap the header using VENUE NOTES A2 in preamble.tex) and templates/chapters/. ICML: copy templates/main_skeleton_icml.tex and templates/chapters_icml/ (renamed to chapters/). Then copy templates/preamble.tex and templates/references.bib, and add the venue kit (ICML: templates/venue/icml/README.md). Rename the main file and the .bib to the kit's stem (iclr2027_conference, neurips_2026, example_paper) [all]. Build once to confirm it compiles.
  5. Source conventions from the first line [all]: one sentence per line, never hard-wrapped; replaced wording kept as a % line next to the new one, never deleted; labels sec:/apdx:/fig:/tab:/eq:/alg: + lowercase hyphenated words; the .bib holds trimmed DBLP entries with short-name keys (Adam, ViT, ImageNet) (latex_source.md, structure.md §3).

Step 1. Outline (write it as % comments at the top of each chapter file [some])

  1. Pair everything (rule of two [all]): two limitations of the prior paradigm, two components, two objectives, and two named twin concepts (e.g. "coverage / precision" for a data-filtering method; "target fit / prior fidelity" for a label-shift correction; "sharpness / calibration" for a forecasting loss). Scope: problems, limitations, components and goals pair; the intro P5 evidence roadmap and the number of RQs follow the experiments (3-4). Write the one-to-one mapping as a % comment: intro limitation (i)/(ii) ↔ method component (i)/(ii) ↔ Methodology subsection titles ↔ RQ/finding that tests each (prose.md §1 rule 1). Introduce the named pair with ONE phrasing per paper: "(termed X)", "By X, we mean ...", or "X, meaning that ..." (prose.md §1 rule 10).
  2. One-sentence key idea: "The key/core idea is to ..." [most]. With two components, split it as "(i) ... and (ii) ..." (e.g. "(i) how much weight each example receives and (ii) how many examples each cluster contributes").
  3. Title: "Verb-ing with " or ": "; sober. Method name: an acronym is expanded with bold initials at first use in the abstract, intro P4 and the Methodology opener [some] (e.g. \textbf{D}ifficulty- and d\textbf{IV}ersity-\textbf{A}ware filtering (DIVA)); a non-acronym name may be bold or plain ([some] each).
  4. One epithet reused throughout (e.g. "cheap and faithful"; "training-free and bounded") [most]. Its two words are the negations of the two drawback labels and the two axes of the standard sentence (contributions.md D2-D3, KR2-KR3); no third word pair on the problem side. While results are pending, pick a property that holds by construction, never effective / accurate (contributions.md C.7). Declare it as % epithet: <phrase> next to the concept table so check_paper.py counts it in the abstract, Introduction, Methodology opener and Conclusion.
  5. RQ list (3-4): performance → mechanism → ablation → extension (experiments.md §2). Each RQ backs one claim.
  6. Notation table: every symbol once, calligraphic text objects, hats for generated quantities (math.md §1).
  7. Running example [most]: pick one concrete example (a real input, question or prompt) that will appear in at least two places: the teaser, the framework figure, and one sentence in the text (prose.md §1 rule 11).
  8. Concept table (concepts.md): list the 2-5 core concepts and ≤6 framing terms as % concept: lines. For each, borrow the established term from the literature and cite it; compose from existing terms if needed; coin only as a last resort (≤2 per paper). Record provenance and definition; one concept = one exact phrase for the whole paper, and every core concept is echoed later (Methodology, Experiments, Conclusion).
  9. Term table (concepts.md §3), before any drafting: one row for every key term (concept entries, every "(termed X)", method, component and metric names, recurring technical nouns): exact phrase | meaning in one plain sentence | verdict (borrowed / composed / plain / coined) | source quote (≤ 15 words) + citation key from the cited papers' own text, same sense | other senses in the field. Write it as % concept: / % term: lines. Prefer the cited papers' terms; coin at most 1-2; a "fixed terms" list never exempts a term (fixed = not reworded, not verified); check senses, not spellings (probe, base, pre-registration, seed noise, screening, share-aware, ablation); one thing = one name; re-audit whenever a term is added.

Step 2. Page budget

  1. Pick the blueprint for the venue (structure.md §5):
    • ICLR, 9 pp [most]: abstract ≤250 words; intro 650-750 (~2 pp); Preliminary 350-550; Methodology 900-1,100; Experiments 850-1,000; Limitations 35-55; Conclusion 80-120. Total ≈3,500-3,700 prose words + 6-8 floats.
    • NeurIPS, 9 pp [some]: ≈3,700-3,850 words.
    • ICML, 8 pp, two columns [venue: ICML]: ≈3,650-3,900 words fill 8.0 pp with about 12 floats (2.6 pp). Most floats are single-column; only method diagrams and the main table are figure*/table* (figures_tables.md §6.1).
  2. List the floats with target sizes: teaser (0.4-0.5 pp one-column; \columnwidth at ICML), framework figure* (0.4-0.5 pp), main table* (0.35-0.5 pp), 2-3 result figures (4-6 column figures at ICML), an optional algorithm (~0.3 pp).
  3. Assign every other item to the appendix now (structure.md §6-7): Related Works, proofs, full setup, extra models/datasets, sweeps, case studies, statements.

Step 3. Drafting order

The order follows the dependencies between sections: the intro restates method notation (P4) and names each experiment subsection (P5), and the abstract and conclusion mirror the intro.

  1. Preliminary (roadmap; task formulation + base paradigm equations; the base-paradigm subsection ends with the limitation the method removes).
  2. Methodology (the type's plan from paper_types.md; roadmap opener; purpose-clause subsection openers; per-term bold run-ins; the last subsection closes the loop; algorithm).
  3. Experiments (RQ list, setup run-ins, tables at final shape, observation blocks). No data yet → experiments.md §7.
  4. Introduction (5 paragraphs, prose.md §3 + the type's P2-P4), with teaser. Write the chain first (introduction.md §0) as a % block of 8-10 one-line claims, and draft only when every link holds: a root-cause sentence that names, neutrally, the one design all prior practice shares (after describing that practice in your own notation, with each weakness planted as a neutral detail); the standard before the weaknesses ("Intuitively, effective X should be A (\ie ...) and B (\ie ...)"), set before the first "However"; each drawback = mechanism + consequence, violating one property; the key idea as the literal negation of the root cause (reuse its content words, flipped); a challenge sentence ("However, ..." with a citation) before any component that does not follow from the key idea; the two axes = the method epithet, repeated verbatim in the proposal, mapping, benefit and closing sentences. P5 is the First/Second/Third evidence roadmap; no contribution list (contributions.md D1). Then check IR1-IR15.
  5. Abstract (prose.md §2), then Limitations and Conclusion (mirror the abstract).
  6. Appendix (Step 5), then Related Works (Appendix A). For each paragraph: open the slot template, look at the matching example in examples.md, draft, check the density norms (prose.md §11) and the sentence rules (§12: one subject per sentence, subject and verb in the first words, the point at the end of the main clause, "we" for method steps and the method name for results), then apply the revision rules (prose.md §9). Wording (prose.md "Use attested expressions"): pick connectives by function from prose.md §4a (preferred tier first; never ref-only or foreign ones) and make every non-first sentence link back (§4b). Take each sentence's wording from assets/lexicon/lexicon.md's section for that rhetorical move, and check any multi-word phrase you are unsure about with python3 tools/ps.py "phrase" -q (queries the shipped phrase index; add --corpus <dir>/corpus.jsonl to check your own papers instead); prefer core and occasional entries, never retired, banned or ref_only ones.

Step 4. Figures and tables

  1. Plots: templates/plot_style.py (apply(), one 600-dpi PNG per panel, save() also writes a PDF, which ICML prefers); palette blue #649CD8/#BFD8EF, pink #FF85A1/#FFDAE2; saturated = ours, tint = baseline (figures_tables.md §3-4). Compare with assets/plots/.
  2. Diagrams in TikZ on templates/housediagram.sty, imitating assets/diagrams/ (figures_tables.md §5): copy templates/diagram_skeleton.tex to figures/src/<name>.tex, set % target-width-in: to the printed width of the float, and place every node relative to another (never at hand-computed coordinates). The package styles carry the house look: white rounded modules with 1.1-pt role outlines (rose trainable, navy frozen), exactly one filled focal module (hero), quoted examples in chips, edgeless lanes, the soft Gaussian shadow, notched heads, dashed amber score flow, solid amber feedback flow, bold sans labels with LaTeX math, a compact legend row; assets/diagram_src/style_sheet.png names every style. Build with python3 scripts/render_diagram.py figures/src/<name>.tex --out-dir figures/framework --check (vector PDF at the printed width + PNG; it fails on Type 3 fonts or a width more than 3 % off the target and prints the clean-palette metrics: white ≥ 0.72, earthy ≤ 0.05, ≤ 3 hue families, mean-sat 0.25-0.75), then open the PNG and fix what you see (wrapped chips, labels on lines, arrowheads off their targets) before including the PDF unscaled. A slide-tool figure is fine if it matches §5.1-5.2.
  3. Tables from templates/tables.tex; choose the main-table variant by the rule in figures_tables.md §7 (T1 / T2 / T4, each [some]). Empty cells are \pend until real numbers exist.
  4. Captions: figures below (title phrase + takeaway; framework captions walk the arrows), tables above.

Step 5. Appendix and statements

Order: A Related Works → B Theoretical Analysis/Methodology → C Implementation Details → D Prompts → E Additional Experiments (mirror the RQs) → F Analysis/Case Study → venue sections (NeurIPS: Broader Impacts, Asset Licenses, then the checklist, also kept at camera-ready). Each top-level section starts with \clearpage [most] and uses an apdx: label; a main-text paragraph that lost material to the appendix ends with a specific Appendix~\ref{apdx:...} pointer. Before the references: ICLR 2027 AI Use Statement (required) + Ethics + Reproducibility statements; ICML \clearpage + Impact Statement (required at submission). ICML appendix: \newpage\appendix\onecolumn [venue: ICML]. Details: venues.md §1.5, §1.7.

Step 6. Compile and page check

  1. If the machine has no LaTeX engine or the build tool is unknown, run bash scripts/check_env.sh first and follow the install command it prints (all options: README.md "Setup"; without a local engine the user can build on Overleaf). Then tectonic -X compile --keep-intermediates main.tex (or latexmk -pdf), with network access for the first build. Fix every error, undefined reference and "Missing character". If tectonic times out downloading a font (some fonts, e.g. cmti9/cmsy5, can time out), run scripts/tectonic_fetch.sh <dir> --missing-pfb and add -Z search-path=<dir> (venues.md §2.3).
  2. python3 scripts/check_paper.py <paper_dir> --stage submission (from this skill folder; the venue and kit year are detected from the main file, --venue overrides, --limit N sets the current call's limit, --vocab PATH adds the rare-word check against your own word list). It prints prose words per chapter against the blueprint, the page of \label{main-text-end} (placed after the Conclusion paragraph), venue checks (statements, checklist, anonymity, ICML float rules), the concept table and echo check, the == terms block (missing citation keys or quotes in % concept: / % term: rows, coined rows above 2, "(termed X)" without a row, sense-clash words such as probe / pre-registered / seed noise), the vocabulary check, cohesion notes (connective share and density per chapter, double lock, approximate isolation, ", so" rate, "(termed" count), lint hits, the == intro & contributions block (contribution-list phrasing; "novel" / "first" / "To the best of our knowledge" in abstract, intro and conclusion; "Different from them" / "Our work differs" / "is not our contribution"; sentence-initial "Ideally,"; "we argue"; intro and abstract paragraphs ending on a reservation; the % epithet: count in abstract, Introduction, Methodology opener and Conclusion; strong result verbs while tables hold -- / \pend cells), and the == habits warnings with line numbers.
  3. The main text must end exactly on the limit page, and that page must be full [all]: never a half-empty last page, never an overflow. Too long → move material to the appendix, tighten floats with negative \vspace inside floats only, toggle % \clearpage. Too short → bring back appendix material (a second model size, a mechanism figure). Never \small body text, shrink fonts, or put negative \vspace at headings.
  4. Look at the rendered pages. The teaser should sit at the top of p2 in one-column venues and at the top of the right column on p1 at ICML. The skeletons force this with \suppressfloats[t] [skill choice] (a template helper, not source practice). Also check: no floats overlapping, no table text larger than body text (\resizebox enlarges narrow tables; use \shrinktofit), and at ICML no figure* stranded after the page it belongs to.

Step 7. Self-review

Go through the checklist below and the checklists at the end of math.md, experiments.md and figures_tables.md; for the Introduction, Related Works and every contribution sentence, run the rule checks of introduction.md §6 (IR1-IR15) and contributions.md E (KR1-KR14). Then run the revision rules (prose.md §9) over every paragraph. Read the == cohesion, == habits and == sentences blocks of check_paper.py (rewrite every listed sentence or confirm it matches the source shape, prose.md §12d): fix each habit warning with the listed alternative (or confirm it is a legitimate sense), bring chapters below the connective targets up with bridges and double locks, and label the Experiments sentences C/A/L/S/N by hand if the approximate isolation rate is high. Finally grep for leftovers: \pend, [Expected], \figplaceholder, <...>, Placeholder-.

Style rules checklist

Venue (references/venues.md)

  • The current year's call, author instructions and kit were fetched at Step 0 and re-checked in the deadline week; the limits and required statements are recorded at the top of the main file.
  • Required statements are present and placed per the kit: ICLR 2027 AI Use Statement before the references; NeurIPS checklist after the appendix, every "Yes" backed by text; ICML Impact Statement (unnumbered, before the references). At submission: no acknowledgements, anonymous code link, self-citations in the third person, no \iclrfinalcopy/[final]/[accepted].
  • Citation commands match the venue: ICLR \citep{} + \Eqref{}; NeurIPS \cite{} numeric + "Equation~\ref{}"; ICML \cite{} or \citep{} (the sty maps \cite to \citep; one argument only; multiple keys in chronological order) + "Equation~\ref{}".
  • ICML [venue: ICML]: [T1]{fontenc} loaded (otherwise Latin Modern fallback); no wrapfigure; no \textwidth in column floats; figure*/table* only for method diagrams and the main table, sourced one page before they print; displays broken with aligned to fit 3.25 in; \printAffiliationsAndNotice{} called.

Concepts (references/concepts.md)

  • Concept table exists (main file, after \input{preamble}): % concept: phrase | core/framing | borrowed: citekey "quote" / composed: parts with citekeys / coined | definition; 2-5 core (max 6) + ≤6 framing.
  • Term table (concepts.md §3): every key term (concepts, each "(termed X)", method / component / metric names, recurring technical nouns) has a % concept: or % term: row with a citation key and a quote (≤ 15 words) read in the cited paper's own text, same sense; ≤ 2 coined rows; no term exempted as "fixed"; no sense-clash word (probe/probing outside linear probes, pre-registered, seed noise, base update, share-aware); re-audited after every rename; check_paper.py == terms shows no missing source.
  • Borrowed before coined: every marked term that exists in the literature uses its established wording and is cited at first use; ≤2 coined terms, introduced once with "(termed X)" and built from existing vocabulary (e.g. budget score, built from the established "sampling budget"). "(termed" count ≤ number of core concepts; the named pair uses one of "(termed X)" / "By X, we mean ..." / "X, meaning that ..." for the whole paper.
  • Marked only at first mention in the Introduction and at the first formal definition (most marks, over three quarters, are in the Introduction; none in abstract, experiments, conclusion, appendix, captions, headings); 8-15 marks (cap 20); no bold/italics on top of the colour.
  • Consistency: the same phrase, abbreviation and symbol for a concept everywhere (text, figures, captions, section titles); no synonyms after first use.
  • First-use explanations (prose.md "Explain at first use; no jargon"): no jargon a cold reader cannot parse; every non-obvious term gets "(\ie ...)" or "(termed X)" at its first occurrence (abstract and Introduction each), then stays unchanged. The explanation sits inside the sentence that carries the link (parenthesis or final clause), never as a separate sentence between a claim and its continuation.
  • Plain words (prose.md "Plain words, no coinage"): no coined compound or rare word outside the core concepts; hyphenated compounds ≤41 per 1,000 main-text words (source papers: about 14-41, median about 27); review the lint's rare-word list when a --vocab file is given.
  • Echo: every core concept reappears, same phrase and unmarked, in Methodology headings/text, the Experiments RQ text or findings, and the Conclusion.
  • One colour (orange \concept{}) ON while drafting/reviewing; \showconceptsfalse for the submitted PDF (black, as in every source paper).

Structure

  • Skeleton: Abstract → Introduction (teaser) → Preliminary → Methodology → Experiments → Limitations → Conclusion; Related Works in Appendix A [most].
  • No contribution bullet list, no "rest of the paper is organized" paragraph [all] (drafts that had them deleted them).
  • The Methodology opens with a roadmap naming its subsections: "We first ... After that, ... Then, ... Finally, ..." [all]; so does the Preliminary [current]. The Introduction has none; the Experiments section opens with the RQ list.
  • The base-paradigm subsection ends with the limitation the Methodology removes [all] (a later background subsection, if any, follows it).
  • One-to-one mapping [most]: intro limitation (i)/(ii) ↔ method component (i)/(ii) ↔ Methodology subsection titles ↔ RQ/finding, same order and same phrase at every link; the P5 roadmap and the RQ count follow the experiments, not the rule of two.
  • The Methodology ends with the subsection that closes the loop: the assembled method, the joint optimization with its algorithm, or inference-time behaviour.
  • Experiment subsections: prefer "(RQk) Noun Phrase" [current]; the suffix form [older] is not wrong; one form per paper. Intro P5 has one "(\cf Section~\ref{})" sentence per subsection, in the same order [current].
  • Limitations: its own numbered section before the Conclusion [most]; 2-3 sentences, two items ("On the one hand ... On the other hand ..." is preferred [current]). Conclusion: one paragraph of 80-140 words that mirrors the abstract [all].
  • Every equation needed to re-implement the method is in the main text; proofs, full setup and extra results go to the appendix, each with a specific pointer [all].

Introduction logic (references/introduction.md §6)

  • The chain was written first as a % block (introduction.md §0) and still matches the prose.
  • Root cause (IR1): one neutral sentence names the design all prior practice shares, before or as the first sentence of the criticism [most]; each drawback's mechanism reuses a noun phrase from it, and each drawback has a planted source phrase in the prior-work description (IR2) [most].
  • Standard first (IR3): the two axes are set ("Intuitively, effective X should be A (\ie ...) and B (\ie ...)") before the first "However" [some]; they equal the method epithet and recur verbatim ≥ 3 more times [all].
  • Each drawback = mechanism + causal connective + consequence (IR4) [all] in method papers; amplifiers carry a number, figure or citation; only drawbacks the method fixes (IR5) [most], one numbered component per drawback.
  • Key idea = literal negation of the root cause, ≥ 2 content words reused and flipped (IR6) [all]; every modifier or component is motivated earlier or by a challenge sentence with a citation (IR7).
  • One centre (IR8): every paragraph's last sentence lands on the two axes; theory supports a component inside the method paragraph; no method paragraph or intro ending on however / unclear / "whether ... any benefit".
  • Prior work (IR10, introduction.md §4): own notation, concede before turning, blame the design not the authors, at most one named modified method [all], closest competitors in a citation cluster; no "we argue", no sentence-initial "Ideally," (IR14).
  • Related Works closers (IR12): each paragraph ends with "While effective, such X ..." in the intro's drawback words or "... remains largely underexplored"; only the last one writes "To address these gaps, we propose ..." with the epithet and component names [most]; never "Different from them", "Our work differs from them", "is not our contribution" (absent from the house finals).
  • A rewritten intro is not longer than the old one by the same counter (IR13); numbers in benefit sentences are attributed to the design that produces them, theory states its premise (IR15).

Contribution emphasis (references/contributions.md E)

  • No contribution list [all]; no "novel", self-referential "first", "for the first time", "To the best of our knowledge" in abstract, intro and conclusion (KR1) [all].
  • One epithet, verbatim in abstract, intro, Methodology opener and Conclusion (KR2) [most], declared as % epithet:; its halves negate the two drawback labels (KR3) [most].
  • Component names unchanged in the intro, method subsection titles (" with "), roadmap, RQ labels and conclusion (KR4); each drawback label recurs verbatim in a method opener, an RQ, the conclusion and a RW closer (KR5) [most].
  • Novelty by contrast with a named shared design, each difference as "rather than ", scope exactly as wide as the works that do it (KR6) [all].
  • Verbs by evidence tier (KR7, A.2); with --/\pend cells left, no demonstrate / validate / verify / prove / consistently / significantly for own results and no "are tested" (C.7).
  • Numbers only in the evidence layer (KR8) [all]; outperform bounded with one qualifier in abstract, intro, bold sentences and conclusion (KR9).
  • Reservations are RQ premises, not closers (KR10); the Conclusion is the abstract in the past tense, same order, no qualifier dropped (KR11) [all]; Limitations = one cost + one scope, checked against every slogan and label (KR12).
  • Teaser caption states the contrast with paired verbs and names the components (KR13) [most]; gains attributed to named mechanisms, borrowed parts credited in the same sentence (KR14, D13, D15) [all].

Prose

  • Abstract: one paragraph, no numbers, citations or dataset names [all]; at most 250 words [most]. Aim for 120-200 at ICML (4-6 sentences) and NeurIPS [current].
  • Rule of two: problems, limitations, components, objectives as "(i) ... and (ii) ..." pairs [all].
  • Running example [most]: the same concrete example appears in at least two places (teaser, framework figure, one sentence in the text).
  • Every problem sentence is followed by a remedy sentence ("To this end / To address this, we ...") [all].
  • Second sentence of a paragraph: "Specifically, ..." [all]. Definitions and examples go in parentheses: (\ie ...), (\eg ...) [all].
  • Cohesion (prose.md §4b): every non-first sentence links back (connective, anaphor, repeated noun or structure); ≤ 5% unlinked (Intro/Method ≤ 3%); connective-led share 45-55% (Intro ≥ 55, Method ≥ 45, Experiments ≥ 35), ~2.2 per 100 words; connective + this/these/such at every problem → solution turn; ≥ 3 shell nouns ("This crowding ...") per Introduction; every paragraph opener links back; RQ subsections open with a bridge and end with a conclusion, not an appendix pointer; "It/Its/This" only for the previous subject; ", so" + clause ≤ 1 per 1k words.
  • Connectives chosen from prose.md §4a: preferred tier first; no older-style (Hence, Thus, Towards this end, As such, In particular, Clearly, In a nutshell, One possible reason ...) and no foreign ones (Ideally, Nor, likewise, is not free, empirical question, whenever).
  • check_paper.py == habits warnings fixed or justified (prose.md "Habits to avoid"); unsure phrases checked with tools/ps.py (shipped index, or your own corpus).
  • Page-fit cuts kept every link-carrying clause; the sentence after each cut re-checked for a link (prose.md §9 rule 17).
  • Emphasis is bold only [all]: bold run-in heads \textbf{Head.} (no \paragraph [most]), bold full-sentence findings after "We have the following observations:" [most], at most one bold-italic research question [some]. Italics only for a term at its definition [some]; no takeaway boxes.
  • Tense: present for the method, experiments and findings in the main text [all]; the Conclusion in past [most] or present [some], one tense per paragraph (prose.md §10).
  • Density: sentences of median 20 words, ≤7% of 40+ words [all]; split any clause of 40+ words [some]; bold findings 8-20 words; sentences per paragraph by section in prose.md §11.
  • Sentence structure (prose.md §12) [all]: section means within range (Abs/Intro ≤ 26, Method ≤ 27, Exp ≤ 21 words), ≤ 1 sentence of 45+ words per section, a long sentence followed by a short one; one subject per sentence (two main clauses ≤ 15%), ≤ 10 words before the subject, relative clauses at the end, ≤ 1 parenthetical per sentence; "Since" openers ≤ 2%, This/These/Such openers ≥ 5% in Intro and Experiments.
  • Subject consistency (prose.md §12b) [all]: we = method steps, the method name = results, one component = its paragraph, This/These = handover; every switch is derived, linear, an anaphor or a parallel contrast opener; no prior-work sentence between two sentences about our method.
  • Emphasis (prose.md §12c) [most]: the point is in the main clause and at the end; our side of a contrast in the main clause, the baseline in a final "while" clause; no result inside a "Since" clause, a parenthesis or a final "which" clause; pointer endings ≤ 12%, no stacked "(termed)" + citation + pointer tails; no fragments or clefts.
  • Calibrated claims: own results firm but bounded ("generally outperforms", "on most of the tasks", "up to"); interpretations hedged ("may", "is likely due to"). Never "always", "state-of-the-art", "SOTA" (absent from [current] papers). About 3 hedges and 2 boosters per 1k words (prose.md §4b rule 9).
  • Banned: delve, elucidate, underpin, furnish, Surprisingly, sheds light, Furthermore, Notably, To be specific (prose.md §8). Avoid a bare "Here," that adds nothing [some]; "Here, X denotes ..." stays fine [some].
  • Precise verbs over verb + adverb [some]: endow, mitigate, impede are used widely [most]; augment, amplify [some].
  • Terminology unified across text, figures and symbols (a frequent revision target [some]).
  • Method name typed literally (no macro); references typed by hand with a tie: "Figure~\ref{}", "Table~\ref{}", "Section~\ref{}", "Appendix~\ref{}", "Equation~\ref{}"; no cleveref [all]. The sources mostly type a plain space; the tie is a [fix] that only prevents a line break before the number.

Math

  • Calligraphic sets and text objects [all], hats on generated quantities [most], bold upright vectors and matrices [most], \mathbb{I}. Skill choices: labels in \text{} superscripts [skill choice] (some sources use italic subscripts) and optional \triangleq [some] (math.md §1).
  • Every display has a colon lead-in, punctuation inside [most], a "where ..." clause and one so-what sentence.
  • Formal results only when real [some]. One numbering scheme per paper: plain counters "Definition 1", "Lemma 1" without a headline theorem [some]; restatable numbered within the section, "Theorem 3.1", with one [some]. Motivate, state self-contained, bold the key relation, sketch or pointer, full proof in the appendix.

Experiments

  • Opener: one sentence that states the section evaluates X and leads into the questions; use the skeleton opener "We evaluate X through the following research questions:" + 3-4 RQ bullets. An appendix pointer in the opener [some] and topic labels [some] are optional.
  • Setup run-ins: \textbf{Models.} (\textbf{LLMs.} when the backbones are language models; ≥2 sizes or families [all] where models are compared), then the data paragraph (\textbf{Training and evaluation.}, \textbf{Evaluation Benchmarks.}, or one head per benchmark family; [some] each), under \subsection{Experiment Setup} or directly under the section ([some] each).
  • \textbf{Baselines.} paragraph when there are ≥2 baseline families [some]; a single natural baseline is named inside the performance RQ with the shared starting checkpoint [some].
  • Finding = bold full-sentence claim + evidence + one interpretive sentence [all]. Optional moves: attribution "We attribute ..." [most], contrast "In contrast, ..." [most], closer "These results ..." [most] or "Such results ..." [some]. Point to the appendix wherever material moved (a pointer in every RQ subsection is [some]).
  • Ablations: "X (w/o C)" variants [some]; optional randomised-signal control [some] and a sweep that justifies the default [some]. NeurIPS: \textbf{Statistical significance.} and compute disclosure answer checklist items [venue: NeurIPS].
  • No fabricated numbers: pending cells are \pend, expected findings marked \draft{[Expected]} until data exist; predictions fixed in advance use the one block template of experiments.md §7 step 5 (intro sentence naming the source → bold prediction → "If ..., then ... should ..." → "The prediction is rejected if ..."), ≤ 2 "is expected to" per section.

Figures and tables

  • One PNG per panel, bold Times, four-sided box, no plot titles, unit in parentheses [all]; blue/pink families, saturated = ours [current]. Current look: 600 dpi and a dashed y-only grid [current]. Axis-label case consistent within the paper (Title Case [some]; sentence case [some]). #595959 spines and inward ticks [some].
  • Diagrams (figures_tables.md §5): built from a TikZ source on housediagram.sty by render_diagram.py --check at the printed width and included unscaled (or a slide-tool figure that matches §5.1-5.2); exactly one filled focal module; one colour = one role (rose trainable, navy fixed, amber score / feedback, teal / coral marks); no free-floating formula; the running example inside; a compact legend row or none; the measured details of §5.2 (1.1-pt strokes, 1.45-pt frames, Gaussian shadow on free-standing objects, notched 5.2-pt heads, r 4.2 elbows, 3.3 / 3.3 dashes, square dots) come with the styles [current]; the PNG checked by eye.
  • Captions: figures below (title phrase + takeaway), tables above ([all] in the main text; required by all three kits).
  • Main table: table* [all], booktabs + | after the row labels [most], width by scaling only, bold header, bold best, \midrule\midrule between scales [current]. Variant by figures_tables.md §7, each [some]: scale-grouped, shaded row, no Avg / per backbone, Avg, bold "+ X (Ours)" / tinted quality + cost rows. ICML: no gray shading behind text.
  • Arrows in headers as $\uparrow$/$\downarrow$ [fix]: the sources typed the Unicode ↑, which drops out under XeLaTeX with Times, or left the arrows out.
  • Floats [t] with negative \vspace inside the float; the main text fills the page limit exactly [all].

Source (LaTeX) (references/latex_source.md)

  • Files: main .tex and .bib keep the kit's stem [all]. Use one chapters/N_name.tex per top-level section, each opening with its \section{} \label{} and pulled in by \input{chapters/N_name} from the main file only: no \include, no nested \input, no separate float files [all]. Rename a chapter file when its content changes [fix] (the sources often kept stale names).
  • Labels: <type>:<lowercase-hyphenated> [most]; sec: main text, apdx: appendix [most]; panels fig:<name>-a/-b [some]; appendix floats fig:apdx-<name> [some]; stock names fig:teaser-figure, fig:main-figure, tab:main-table, sec:main-performance, apdx:impl-details, apdx:case-study. The \label goes on the heading line, right after \caption{} [all], and on the \begin{equation} line [most]. No duplicate labels and no uncited main-text float (check_paper.py flags duplicates).
  • Figure files [all]: \includegraphics{figures/<topic>/<file>.png} from the project root, with the extension, no \graphicspath. Diagrams go in figures/framework/, plots in figures/exp/<analysis>/<content>_<model>.png [current]. Never use numbered names. Superseded files are renamed (*_old), not deleted.
  • A float is placed inline in the subsection of its first \ref (nearly always), right before the paragraph that cites it or right after it when that paragraph ends the subsection. The teaser is the first thing after \section{Introduction} [most].
  • Macros: ICLR puts the paper's own macro block in math_commands.tex [venue: ICLR]; NeurIPS/ICML in the main preamble [all]; the templates keep it in preamble.tex. Chapter files define nothing global.
  • Revision trail [all]: old sentences, paragraphs, floats, headings and colour values stay as % lines. One co-author macro per person, named by initials (\coA{}, \coB{} in the template; rename them); \concept{} marks coined terms [most]; all are black at submission and the wrappers stay. At ICLR, discussion-phase changes go in \rebuttal{} or \revise{} [venue: ICLR].
  • Bibliography [all]: DBLP exports trimmed to author/title/venue/volume/year, venue as the acronym, arXiv as journal = {CoRR}, short-name keys grouped under % Topic headers; scripts/bib_to_dblp.py converts a verified bib to this form (latex_source.md §8). No duplicate keys, and \cite keys match the entries' case. Strip comments with a cleaner tool before an arXiv source upload.

Templates

FileWhat it is
templates/preamble.texConsolidated packages, macros (\ie, \eg, \cf, math kit, theorem set with both numbering schemes: plain counters and a within-section restatable thm), colours (table shading, change superscripts, roles, plot palette), prompt/case boxes, algorithm phase shading, \pend, \ablation, \shrinktofit, \figplaceholder; venue header notes for ICLR 2027 / NeurIPS 2026 / ICML 2026 at the top
templates/main_skeleton.texICLR 2027 main file with the house chapters/ \input layout (RW commented out; AI Use + Ethics + Reproducibility statements before the bibliography; appendix after it). NeurIPS: swap the header (VENUE NOTES A2)
templates/chapters/*.texType-neutral one-column section scaffolds with slot comments (teaser, framework figure, Definition/Lemma/restatable theorem, shaded algorithm, empty main table, appendix sections, ICLR statements); per-type wording in references/paper_types.md
templates/main_skeleton_icml.texICML 2026 main file in the two-column house layout [venue: ICML]: blind/preprint/accepted lines, [T1]{fontenc}, \twocolumn[...] title block, \printAffiliationsAndNotice{}, \clearpage + Impact Statement, \onecolumn appendix
templates/chapters_icml/*.texTwo-column scaffolds with every float type of the house layout at its width (column teaser, prompt/text box, framework figure* 0.9, shaded column algorithm, figure* 1.0, T4 table* with \pend, column table and plots, subfigure pair, aligned objective) and post-label \vspace trims (none on a float stacked directly above another float); no pre-caption trims (ICML needs ≥0.1 in around captions). 8_appendix.tex holds every appendix section. Copy to chapters/
templates/venue/icml/README.mdWhere to get the ICML kit (not bundled: icml2026.sty has no license statement) and how to swap the year
templates/tables.texT1 scale-grouped main results, T2 per-backbone "+ X (Ours)" with Avg, T3 before/after superscripts, T4 quality + cost rows, T5 ablation, T6 sweep, T7 hyperparameters, T8 dataset statistics
templates/tables_preview.texHarness that renders every table sample on its own page
templates/plot_style.pyrcParams, palette (set_palette() swaps the hues, keeping the roles), sizes, legend/twin-axis/value-label/rolling-mean helpers, save() to 600-dpi PNG + PDF
templates/housediagram.styTikZ style package for diagrams: the house palette and type scale, module styles (trainable, frozen, neutral, state, hero, hero rose, hero gradient, ghost), items (input square, chip, token and their selected / dropped forms, tagged=), containers (\hdlane, \hdcard, \hdperforated, \hdframe, \hdgroup, \hdstack), connectors (data flow, role flow, score flow, feedback flow, ghost flow), the vector Gaussian shadow, marks (good, bad, frozen), callouts (callout box + \hdtail) and legend keys (legend row, key box, key pair, key arrow, key mark); [serif] option for Times labels
templates/diagram_skeleton.texstarter diagram (example lane, two components, focal module, model, feedback loop, headers, legend row) with the build command; copy to figures/src/
templates/references.bibthe key convention + three placeholder entries so the skeletons compile; replace
scripts/check_paper.pyword budget per chapter and stage, main-text-end page vs the venue/stage limit, venue checks (kit year, statements, checklist, anonymity, ICML float rules), concepts, term provenance and sense-clash words (== terms), vocabulary (--vocab), cohesion notes (soft targets), style lint incl. duplicate labels (venues.md §4), intro & contribution checks (introduction.md IR, contributions.md KR; % epithet: count), and habit warnings with line numbers (--habits-max N)
assets/lexicon/lexicon.md, assets/lexicon/lexicon.jsonthe shipped 1,272-entry phrase bank (readable / machine-readable), by writing step, each entry with function, per-group counts, tier and an example (tools/README.md)
tools/phrase_index.json.gzthe shipped 1-5 word n-gram counts (no sentence text) behind ps.py's default (no---corpus) mode, over the same house/refA/refB papers as the lexicon
tools/build_corpus.pybuilds a sentence corpus (corpus.jsonl) and a word list (vocab.txt) from YOUR main .tex files (--mine, optional --reference, or --config), into a workspace folder outside the skill
tools/ps.py, tools/ngrams.pyps.py phrase search: queries the shipped phrase index by default, or your own --corpus (per-paper counts, rate, examples and tier; --draft adds your current draft); ngrams.py lists recurring n-grams and openers of your own corpus (tools/README.md)
scripts/bib_to_dblp.pyrewrites a verified .bib in the DBLP-condensed style: short venue names, arXiv as journal = {CoRR}, full author lists (completed from the arXiv API), "First Last" names, acronym-only title braces (latex_source.md §8)
scripts/check_env.shenvironment self-check: Python, matplotlib/numpy, PyMuPDF, LaTeX engine, Times font (plots), test builds of templates/env_test.tex and templates/diagram_skeleton.tex; prints install commands (README.md "Setup")
scripts/render_diagram.pybuilds TikZ diagram sources into a vector PDF at the printed width + a PNG preview: tectonic / latexmk / pdflatex with templates/ on the search path, warnings that change the output, width vs % target-width-in: (fails beyond 3 %), Type 3 fonts and raster images, --check for the clean-palette metrics (PyMuPDF)
scripts/tectonic_fetch.shfetches bundle files (fonts) in small byte ranges when tectonic's downloader times out
scripts/colour_stats.pycleanliness metrics of a diagram PNG: white share, earthy share, hue families, mean saturation of the coloured pixels, with the targets of figures_tables.md §5.1 (numpy, plus PyMuPDF or matplotlib to read the PNG; render_diagram.py --check calls it)
assets/diagrams/, assets/plots/synthetic reference diagrams (two teasers, framework pipeline, key-idea loop, group-quota mechanism) on one invented method and one running example, each a printed-size vector PDF + 300-dpi PNG, and 1000-px PNG plots, all drawn from invented data; they illustrate the conventions, not real results
assets/make_assets.pyregenerates every file in assets/: the plots with templates/plot_style.py, the diagrams from their TikZ sources in assets/diagram_src/ through scripts/render_diagram.py --check (on templates/housediagram.sty, which implements the clean palette of figures_tables.md §5.1 and the measured details of §5.2), and diagram_src/style_sheet.png

Build check (tectonic 0.17 with network access; re-run it after editing any template). Pass criteria for all three builds: 0 LaTeX warnings, 0 overfull boxes and 0 missing characters.

  • python3 assets/make_assets.py: the five diagrams and style_sheet.tex (and templates/diagram_skeleton.tex with render_diagram.py) build with 0 warnings, 0 Type 3 fonts and 0 raster images, each within 3 % of its target width (most within 1 %), and all five diagrams pass the clean-palette targets.
  • main_skeleton.tex + chapters/ against the official iclr2027_conference.sty: about 12 pp, Times (Nimbus) fonts, teaser at the top of p2, AI Use Statement before the references.
  • The NeurIPS header swap (VENUE NOTES A2) against the official neurips_2026.sty [main]: about 11 pp.
  • main_skeleton_icml.tex + chapters_icml/ against the official icml2026.sty: about 10-11 pp in blind mode, with the line-number ruler; [accepted] and [preprint] also build. The fonts are Nimbus Type 1, with no Latin Modern fallback. Only the template's two hyperref notices remain (the \\ in the title and an empty anchor).

check_paper.py runs for all three venues. On the scaffolds it finds main-text-end on p5 (one column) and p4 (ICML), since the scaffolds are short, and 0 venue hits for ICLR and ICML; the == concepts section reads the skeleton's placeholder concept table, finds both core concepts marked in the Introduction and echoed in the Methodology, Experiments and Conclusion, == terms reads the placeholder rows (3 concept + 1 term) with 0 warnings, and == vocabulary reports the compound rate (the rare-word list runs only with --vocab). For the NeurIPS swap it reports 1 hit (no checklist) plus the \citep lints; both disappear after the conversion steps in venues.md §3.

Skills relacionados