review-paper

Review a scientific paper against the CompLING house-style rules below. The argument is optional: it may be a path to a LaTeX file or a directory of .tex files, optionally followed by additional instructions.

Parse the argument string:

• If the argument is empty, use the current working directory as the project root; everything counts as additional instructions (empty here).
• Otherwise, the first token is the path; everything after is additional instructions (may be empty). If the first token does not look like a path (does not exist as a file or directory), treat the entire argument as additional instructions and use the current working directory as the project root.

Steps

1. Locate source files. If the path is a directory (including the current working directory when no argument was given), find all .tex files in it, then identify the main file (the one containing \documentclass and \begin{document}) and follow \input{...} / \include{...} references from there. If the path is a file, start from that file and follow \input{...} references to any other .tex files in the project.
2. Scan each file for violations of the rules in the sections below. For each violation, record:
   • File and line number (use file.tex:LINE format).
   • A short description of the issue.
   • The suggested fix.
3. Group the findings by category (Writing, Typesetting, Citations, Math, etc.) and present them to the user as a summary.
4. Annotate the source with TODO notes. For each identified content issue, insert a \fredaauto{...} comment at or just before the offending location in the .tex file (see "Inline annotations" below for setup). Keep the comment text short — one sentence stating the issue and the suggested fix.
5. If additional instructions were provided in the argument, follow them after the review.
6. Verify citations. If the project has any .bib file(s), run the reference checker from this skill's scripts/ directory: uv run <skill-dir>/scripts/check_reference.py <bib-path-or-project-dir>. The script queries CrossRef, Semantic Scholar, and OpenAlex and emits a grouped report (verified / suspicious / not_found / skipped). Add its findings under the Citations category of the summary: cite bibfile.bib:LINE for each flagged entry and note the specific reason (author divergence, year mismatch, generic title, malformed DOI, etc.). Do not insert \fredaauto{...} annotations on .bib files — bib findings belong in the summary report only. If the script cannot reach the network (all three APIs fail), re-run with --no-network to still report local red flags, and explicitly note that the existence check was skipped.
7. Check logical consistency and clarity. After all format, typesetting, and citation checks are complete, do a dedicated pass over the prose focused on content-level problems. Look for:
   • Contradictions. Statements in the paper that disagree on facts, numbers, model sizes, dataset sizes, hyperparameters, baselines, results, or claims (e.g., the abstract reports +2.3 points but Table 2 shows +1.9; Section 3 says "we use 5 seeds" but Section 4 reports "averaged over 3 runs"; the introduction claims "first to do X" but related work cites a prior X).
   • Terminology drift. The same concept referred to by different names without explanation (e.g., our model, the proposed method, the framework, our system used interchangeably; token, word, subword mixed within one paragraph). Flag only when the drift is likely to confuse the reader.
   • Ambiguous referents. Pronouns (it, this, they, that, these) whose antecedent is unclear, or demonstratives without a clear noun (this approach, these issues) when multiple candidates are in scope.
   • Unclear or underspecified claims. Vague quantifiers (significantly, substantially, many, often, generally) used where a concrete number, dataset, or condition should appear; load-bearing passive constructions that hide the agent (it was shown that) when specifying the agent matters; undefined symbols or terms introduced without a prior definition.
   • Dangling references. \ref{...} / \cref{...} / \cite{...} targets that do not resolve (missing \label{...}, typos, stale labels), references to sections/figures/tables/equations that do not exist or were renumbered, and phrases like "as we will see in Section 4" when the referenced section does not cover that material.
   • Internal inconsistency in procedures / experiments. Method descriptions whose steps, equations, or dependencies do not line up (e.g., a variable used in Eq. 3 but introduced only in Eq. 5; an algorithm that reads an input never defined; a training recipe whose hyperparameters conflict with Table 1).
   • Abstract / introduction / conclusion misalignment. Contributions listed in one place but absent in another; the conclusion claims a result the experiments section does not actually report; the abstract over- or under-sells what the paper delivers.
   • Figure / table ↔ text mismatch. The text describes a trend that the figure or table does not show (wrong direction, wrong column, wrong condition); a table caption describes different columns than the table has; a figure is cited from the text but never appears, or appears but is never cited. For each finding, record file.tex:LINE (or file.tex:LINE-LINE for a span), a one-sentence description of the problem, and a concrete suggested fix (or two options if the correct resolution depends on author intent). If a finding can be localized to a specific line, insert a \fredaauto{...} annotation following the same rules as in "Inline annotations" above (including the compile-verification step). If the issue spans the whole paper or cannot be localized (e.g., a claim in the abstract contradicting a result in Section 5), leave it in the summary report only. Add these under a Consistency & clarity category in the final report. Where the correct fix depends on author intent you cannot infer from the text alone, mark the item as needs author input instead of proposing a specific rewrite.
8. Save the issue report (final step). The skill must always end by writing a consolidated issue report to review-report.md at the project root (the directory resolved in the argument-parsing step), in the format described in "Output format" below, even if no issues were found (in which case, write an empty report with a one-line "no issues found" summary). Overwrite any existing review-report.md at that path. After saving, point the user to the file. This is the skill's primary deliverable — do not end the turn without it.

Inline annotations

Annotations use the todonotes package. Before inserting any \fredaauto{...} calls, verify the main file's preamble contains the required setup. If any of it is missing, add it to the preamble (right after other \usepackage{...} lines).

Required preamble:

\usepackage{todonotes}
\usepackage{multirow}
\usepackage{array}
\usepackage{tcolorbox}
\newcommand{\note}[4][]{\todo[author=#2,color=#3,size=\scriptsize,fancyline,caption={},#1]{#4}}
\definecolor{tticblue}{RGB}{0, 94, 184}
\newcommand{\fredaauto}[2][]{\note[#1]{Freda via Erich Review Bot}{tticblue!20}{#2}}
\newcommand{\Fredaauto}[2][]{\fredaauto[inline,#1]{#2}\noindent}

Rules for inserting annotations:

• Always use the color tticblue!20 (which \fredaauto{...} already does — do not invent other colors).
• Use \fredaauto{...} for short margin notes. Use \Fredaauto{...} for longer inline notes that would not fit in a margin.
• Place the annotation directly before or inside the problem location so it renders near the issue. Do not reflow or otherwise edit the surrounding text.
• Do not add \fredaauto{...} for issues that cannot be localized to a specific line (e.g., overall structure comments) — leave those in the summary report only.
• Inside floating environments (figure, table, figure*, table*), annotations are allowed but must go outside \caption{...}. Use \Fredaauto{...} placed elsewhere in the float (e.g., just before \caption{...} or before \end{figure} / \end{table}) to flag caption-wording issues. Never insert \fredaauto{...} or \Fredaauto{...} inside the \caption{...} argument itself.
• If a \fredaauto / \Fredaauto macro is already defined in the preamble with a different color or author, leave it alone and use it as-is.
• Verify compilation after each annotation. After inserting each \fredaauto{...} or \Fredaauto{...} annotation (including the very first one, which also exercises any preamble additions), run pdflatex -interaction=nonstopmode -halt-on-error <main.tex> from the project root to confirm the document still compiles. If the compile fails, revert that specific annotation (and, for the first-insertion case, any preamble changes that caused the failure) so the file returns to its previous state, then skip that site. Record each skipped annotation in the final report under a Skipped annotations section with file.tex:LINE, the intended note text, and a one-line summary of the compile error. Do not proceed to further annotations until the file is back in a compiling state.

Only flag issues you are confident about. If a rule has judgment-call edge cases (e.g., tense, voice), note the ambiguity instead of insisting on one option.

Rules to check

Structure and project setup

• Hierarchical organization. The main file should split long papers into per-section files included via \input{...}. Flag papers with one giant .tex file only if it is unwieldy (~500+ lines).
• Macros. Recurring strings (model names, method names, dataset names) should use \newcommand. Flag repeated literal occurrences (3+ times) of the same model/method name that could be a macro.
• Booktabs tables. Tables should use \usepackage{booktabs} with \toprule, \midrule, \bottomrule. Flag any \hline in tables. Flag vertical lines in tables (e.g., | in tabular column specs like {l|cc}).
• Figures. Flag \includegraphics calls loading raster formats (.png, .jpg, .jpeg) unless a vectorized alternative is clearly unavailable (e.g., photographs). PDF/SVG is preferred.
• Caption placement. In ACL-style papers, captions belong below figures and tables (i.e., \caption{...} should appear after \includegraphics or after the tabular environment, still inside the float). Flag captions placed above.
• Cross-references. Prefer \cref{...} (from cleveref) over manual Figure~\ref{...} / Table~\ref{...}. Flag inconsistent or manual reference styles.

Writing style

• Tense. Prefer present tense. Past tense is acceptable for dataset construction or experimental procedures. Flag gratuitous past tense in introductions, related work, and analysis.
• Simpler paraphrases. Flag these words and suggest replacements:
  • utilize, leverage → use
  • in order to → to
  • due to the fact that → because
  • a large number of → many
• Passive voice. Flag passive voice that exists solely to avoid we (e.g., it was observed that → we observe that). Passive voice is fine when the agent is genuinely unimportant.
• Contractions. Flag any contractions in the prose: don't → do not, we're → we are, it's → it is, can't → cannot, etc.
• Possessive 's on inanimate nouns is often awkward. Flag the system's implementation → the implementation of the system (but do not be dogmatic — short possessives like the model's output are usually fine).
• Title capitalization. In \title{...} and section titles (\section{...}, \subsection{...}), capitalize every word except articles (a, an, the), prepositions, and coordinating conjunctions (and, or, but). Always capitalize the first and last word.
• Main-text capitalization. In the body, only capitalize named entities and non-obvious abbreviations. Flag title-case phrases like Natural Language Processing in body text (should be natural language processing). Do NOT flag established named architectures like Transformer.
• Common expressions.
  • Flag e.g. and i.e. not followed by a comma: e.g., and i.e., (two commas).
  • Flag etc. used for people (should be et al.) and et al. used for things (should be etc.).
  • Flag state-of-the-art used as a noun or state of the art used as an adjective.
  • a priori is an adjective phrase; flag noun usage.
• Uncountable nouns. Flag many works → much work; researches (as a noun) → research; literatures → literature; evidences → evidence.
• Quotation marks. Flag straight double quotes "..." in LaTeX source — LaTeX needs ...'' (two backticks, then two single quotes).
• Dashes. Flag single hyphens (-) used to connect clauses or names that should use en dash (--) or em dash (---). Examples: Kullback-Leibler should be Kullback--Leibler; a clause break like something - something else should be something---something else (no spaces) or something -- something else (with spaces). Do NOT flag hyphens inside compound words (state-of-the-art, end-to-end).
• Period inside closing quote when a sentence ends with a quotation: "...world." not "...world".
• Footnote placement. \footnote{...} should follow periods, commas, or closing parentheses with no space. Flag footnotes placed before the punctuation.
• a vs. an. Check by the sound of the following word, not the letter. A university, an LSTM, an MLP, a unique. Flag obvious mismatches.

Citations and bibliography

• natbib + cleveref. The preamble should include \usepackage{natbib} (or the conference-provided equivalent) and \usepackage{cleveref}. Flag absence.
• Citation commands. Flag \cite{...} when \citep{...} or \citet{...} would be more appropriate in context.
• Venue acronyms. In BibTeX entries for papers from the following venues, the booktitle or journal field should use the acronym, not the full name: ACL, EMNLP, NAACL, EACL, AACL, COLING, TACL, CL, CVPR, ICCV, ECCV, ICLR, ICML, NeurIPS, AAAI, IJCAI, AISTATS, COLT, UAI, TMLR, JAIR, TPAMI, CogSci. Flag full-name usage for these.
• Capitalization protection in bib. In BibTeX title fields, words that must stay capitalized (model names, proper nouns, acronyms) should be wrapped in braces: {BERT}, {Grad-CAM}. Flag unwrapped proper nouns/acronyms in titles.
• Citation correctness. Spot-check a few entries for obviously wrong year or venue if possible, but do not fabricate corrections — just flag suspicious entries.
• Hallucinated citations. Every .bib entry should resolve to a real work in CrossRef, Semantic Scholar, or OpenAlex. Step 6 runs check_reference.py to perform this verification; surface any not_found or suspicious results in the Citations section of the summary.

Math typesetting

• Equation environments. Flag \begin{equation} and \begin{eqnarray}. Prefer align or align*. Only numbered equations should use align; unreferenced equations should use align*.
• Multi-character sub/superscripts. Flag $p_{surface}$ style — should be $p_\textit{surface}$ (or similar text wrapper). Single-letter subscripts are fine.
• Ellipsis. Flag ... inside math mode; prefer \ldots for lists ($x_1, \ldots, x_n$) and \cdots for operation sequences ($x_1 + \cdots + x_n$).
• Punctuation in display math. Displayed equations that are part of a sentence should have a comma or period. Flag display equations in the middle of a sentence that end with nothing.
• Variable conventions:
  • Scalars: lowercase Latin letters (a, b, c; integers i, j, k, \ell, m, n; RV values x, y, z). Flag use of l (lowercase L) as an index — prefer \ell.
  • Structured values (sequences, trees): lowercase \boldsymbol{...}.
  • Vectors: \mathbf{...} for Latin, \boldsymbol{...} for Greek.
  • Matrices: capital \mathbf{...} for Latin, \boldsymbol{...} for Greek.
  • Vector/matrix elements: scalar typeface (e.g., h_i, w_{i,j}).
  • Sets of structured values: capital \mathcal{...}.
  • Sets of numbers: capital \mathbb{...} (e.g., \mathbb{R}).
  • Random variables: capital X, Y, Z; vector-valued RVs: \boldsymbol{X}. Flag clear violations (e.g., a weight matrix typeset as W instead of \mathbf{W}), but do not be overzealous on short inline uses.
• Tuples. Flag <x, y> for ordered pairs; use (x, y) or \langle x, y \rangle.
• Probability:
  • Conditional bars: flag | in conditional probability; use \mid (e.g., p(y \mid x)).
  • Sums/products: flag \Sigma and \Pi used as sum/product operators; use \sum and \prod.
  • Expectations: prefer \mathbb{E}.
• Named operations. Flag bare text like softmax(x) — should be \mathrm{softmax}(x). Built-in operators like \log, \exp, \sin, \max, \min should use their macros, not \mathrm{...} or plain text.
• Function names. Functions like score(x) should be \textit{score}(x), not score(x) (which renders as s·c·o·r·e).

Output format

Write the review to review-report.md at the project root as a grouped list. For each finding:

[category] file.tex:LINE — <issue>
  fix: <suggested replacement>

If any annotations were reverted because pdflatex failed to compile (see "Verify compilation after each annotation"), include a Skipped annotations section listing each one with file.tex:LINE, the intended note text, and a brief compile-error summary.

End with a short summary: how many issues in each category, and a one-line overall impression.

Rules while reviewing

• The only edits allowed are (a) inserting \fredaauto{...} / \Fredaauto{...} annotations at issue sites, and (b) adding the required todonotes preamble setup if missing. Do not reflow prose, rename macros, or "fix" anything else without explicit user permission.
• Do not commit anything.
• If the paper uses a conference style file (e.g., acl.sty, neurips.sty) that overrides some of these defaults, defer to the conference style.
• If a rule's application is ambiguous in context, note the ambiguity rather than forcing a fix.