Default standards for English prose. Apply at generation and edit time. All rules are peers; no group is higher priority than another.
RULE-01..12 distilled from Strunk & White, Orwell, Pinker, and Gopen & Swan. RULE-13..17 added by Wenhao Chai.
Additions for page-constrained conference papers (NeurIPS/ICML/ICLR 9-page caps and similar). Field-observed from paper-review workflows.
-
RULE-P1 Do not write section or subsection overview paragraphs. Do not open §3 Method with "We propose X, a framework that does (1) ..., (2) ..., (3) ...". Let the section-level figure caption, the algorithm block, and the component subsections carry the overview. The introduction already introduced the method.
-
RULE-P2 Do not use \emph, \textit, or \textbf for emphasis in body prose. Reserve typographic emphasis for at most 1–2 signature concepts per paper, introduced once. For local emphasis, change wording: pick a precise verb, use parallel structure, or restructure the clause. \textbf in body text is almost never justified; reserve for table headers and figure-caption labels.
-
RULE-P3 Researcher voice. Do not open sections or paragraphs with meta-scaffold sentences: "The motivation is empirical:", "In this section, we ...", "To begin, ...". Start with the concrete claim or observation. Do not plain-ify researcher-formal terms (regime, load-bearing, orthogonal, ablation, identifiable); they carry precision. Delete redundant forward and backward references such as (defined in §3.2) or , as discussed above, when the surrounding text already introduces the concept.
-
RULE-P4 Space economy under page caps. Before adding a paragraph, numbered list, or subsection, ask what information is lost if it is cut; if the answer is "only structural scaffold", do not add it. Do not default to a numbered Contributions list in the introduction; let contributions emerge through the narrative. Keep titles to a method name plus 1–2 signature concepts; stack more than two concepts in a title signals either unfocused framing or a missing abstract.
-
RULE-P5 Do not insert blank lines inside a \paragraph{} block: all prose belongs to a single LaTeX paragraph. If a sub-topic feels distinct enough to need a break, give it its own \paragraph{} header instead. Layout breaks around numbered display equations (\begin{equation}) are exempt.
-
RULE-P6 Audit every equation, display and inline, for undefined symbols. Each free variable, custom function (input/output type), index, and custom operator must be defined in prose: as setup before the equation, as unpacking after, or, for inline math, at the equation site itself where definition and use often coincide. State output types when arithmetic depends on them.
-
RULE-P7 Citation discipline. Every \bib entry must correspond to a real, locatable paper: title, authors, year all verified; do not trust LLM-generated entries without checking, since a fabricated reference is citation fraud and damages credibility more than any prose flaw. Use \citet only when the citation is a sentence component, e.g., \citet{tong24} show that ... renders as "Tong et al. (2024) show that ..."; use \citep everywhere else, e.g., "...as shown in prior work \citep{tong24}" renders as "...as shown in prior work (Tong et al., 2024)". Test by removing the citation: if the sentence is still grammatical, use \citep; if it breaks, use \citet. Order multiple citations chronologically. In .bib entries: prefer the published venue over arXiv when both exist; use @inproceedings for conferences and @article for journals or arXiv-only papers; use the 4-letter conference abbreviation, e.g., CVPR or ICML, not the full name; drop page, volume, and number fields; deduplicate entries. Source .bib from Google Scholar, since arXiv and Semantic Scholar exports use inconsistent formatting.
-
RULE-P8 Use the precision the measurement supports: 81.2% over 81.23% in most cases. Pick math-mode ($82.8$) or text-mode (82.8) and hold it across the whole paper, including tables.
-
RULE-P9 Section titles do work. Use the question or finding the section answers, not a generic label: "Where is the signal?" over "Interpretability"; "Grounding agent reviews" over "Applications". Keep to 1–4 words unless the question form needs more. Avoid subsubsections.
-
RULE-P10 The conclusion must extend, not paraphrase, the abstract. Restate key results in one or two lines, then add implication, interpretation, or forward-looking discussion. Naming it "Conclusion and Discussion" reminds you to do the second half.
-
RULE-P11 Use simple present tense as the default voice, and hold it consistently within a piece. Switch tense only when meaning requires it.
-
RULE-P12 Each paragraph carries one idea. The first sentence states the thesis; the rest supports or extends it; the last either concludes or hands off to the next paragraph. After drafting, re-read and split or merge paragraphs so this holds.
-
RULE-P13 Lead with the simplest example that captures the crux of the problem, then generalize. Generality without a grounded toy case is the enemy of understanding. For every parameter or feature in the toy, check whether removing it keeps the problem interesting; if so, remove it before generalizing.
-
RULE-P14 A figure caption must stand alone: state what is plotted, what the axes are, and what the reader is meant to take away. A reader who skims only figures and captions should still get the main empirical claim.
-
RULE-P15 Treat theorems as the paper's public API. State at most 2–3 main theorems up front, informally if needed; a reader applying the result should not need to read the proof. Push technical lemmas and proof details into the appendix or encapsulated lemmas.
-
RULE-P16 When a defined symbol reappears across pages or sections, restate its semantic role at the use site: prefer "if the rank $d$ is small" over "if $d$ is small". This complements RULE-P3, which deletes ceremonial cross-references like (defined in §3.2); together they remove boilerplate while keeping the concept fresh.
-
RULE-P17 A paper introduces at most 1–2 newly coined terms; every additional name is jargon the reader must learn. Do not invent acronyms. Do not collapse a problem and an algorithm into one name (e.g., "K-means" used for both the objective and the heuristic).
-
RULE-P18 Treat a displayed equation as part of its containing sentence: end the equation with a comma if the sentence continues, a period if it terminates. The sentence after a terminal-period equation begins with a capital letter.
Layout and typography rules that depend on rendered output. Trigger only when reviewing the compiled PDF for submission, not during drafting or generation. Skip these when editing .tex source without a fresh build.
Break any rule sooner than write something awkward (Orwell 1946 Rule 6). Rules serve clarity; they are not ends in themselves.
Additions beyond RULE-12 are by Wenhao Chai, distilled from paper-review workflows and field practice across NeurIPS / ICML / ICLR / CVPR / EMNLP and adjacent venues.
