CS Conference Writing Style

Reviewed: 2026-10-01. ACL/ARR, EMNLP, NAACL, CHI, CSCW, UIST, KDD, WWW, and SIGIR differ by track and cycle. The points below describe drafting choices; they do not impose one empirical-paper structure on all CS research.

NLP: ACL, ARR, EMNLP, NAACL

Primary source: https://aclrollingreview.org/cfp

ARR is a reviewing service followed by commitment to a participating publication venue. Its OpenReview reviews are not public by default. Verify the cycle and committing venue separately. Current ARR requires a Limitations section and provides a Responsible NLP checklist. Starting with the October 2026 cycle, review capacity and qualified service-contributor rules also affect submissions. See conferences_formatting.md for the current content-page scope.

A useful empirical NLP paper explains the task, gap, method, evaluation, and error analysis. Theory, surveys, position papers, resource work, and negative results may need different structures; ARR explicitly admits several of these. A short paper should make a focused contribution, not merely omit necessary parts of a long paper.

Evaluation

• Choose evaluations matching the task, language coverage, population, and claims.
• Report dataset versions, licences, train/dev/test splits, preprocessing, and leakage controls, including possible benchmark contamination where relevant.
• Compare appropriate baselines under comparable tuning and compute conditions.
• Define metrics and averaging; aggregate F1, BLEU, or ROUGE does not establish factuality, safety, or utility on its own.
• Report run-to-run variability and use uncertainty methods appropriate to paired items, repeated participants, or other dependencies in the design.

For human evaluation, report recruitment, qualifications, instructions, sampling, compensation, consent/ethics review, and analysis. Choose an agreement measure suited to the annotation scale and number of raters; Cohen's kappa is not a universal statistic for a three-rater ordinal task. Distinguish agreement from validity of the annotations.

Anonymity and citation

Follow the current two-way anonymization rules for papers, supplements, and links. Use third-person self-citations when appropriate; do not remove relevant prior work. ARR's preprint policy and its binding no-preprint choice are distinct. Do not promise an anonymous public repository without checking its contents, metadata, ownership, and the call's link requirements.

HCI: CHI, CSCW, UIST

Primary CHI 2026 sources:

• https://chi2026.acm.org/authors/papers/
• https://chi2026.acm.org/chi-publication-formats/
• https://chi2026.acm.org/guide-to-a-successful-submission/
• https://chi2026.acm.org/papers-review-process/

CHI 2026 paper review uses a single-column anonymized manuscript. The LaTeX configuration is \documentclass[manuscript,review,anonymous]{acmart}; the old sigchi option is not the current papers-review recipe. TAPS produces publication outputs after acceptance. Other HCI tracks and venues may use different formats and review models.

Contribution and method

Describe the contribution to understanding, designing, or evaluating interaction. Empirical, artifact, methodological, theoretical, and other contributions need evidence appropriate to their claims; not every contribution requires the same user-study design or a business-impact metric.

For participant research, document the ethics review, exemption, or applicable institutional process accurately. Do not imply an IRB is the name of every country's review body or invent approval for a study that did not receive it. Report recruitment, relevant demographics, consent, compensation, setting, and analysis, with privacy-aware presentation.

For qualitative work, explain sampling, data collection, analytic method, researcher positioning where relevant, and how evidence supports interpretations. Use actual participant quotations with appropriate consent and anonymization; never fill a placeholder with a fictional quotation presented as study data.

For system/design work, explain the rationale and evaluate claims at their actual scope. A short laboratory evaluation does not establish long-term deployment benefit. State limits and identify which design implications follow from evidence.

Accessibility and author information

Provide meaningful figure descriptions and readable encodings following the current ACM/venue instructions. Verify contribution-statement requirements rather than assuming one mandatory per-author section across all HCI venues. Check the final PDF and HTML proofs generated by TAPS.

Mining and IR: KDD, WWW, SIGIR

KDD 2026 Research Track source: https://kdd2026.kdd.org/research-track-call-for-papers/

Separate Research, Applied Data Science, Datasets and Benchmarks, AI for Sciences, and other tracks. KDD's research review setting uses \documentclass[sigconf,anonymous,review]{acmart}, which is intentionally different from CHI's single-column review setting. Use the actual current call for WWW/SIGIR and for other KDD tracks.

For scalability claims, report problem size, data properties, complexity assumptions, hardware, software, runtime, and peak memory. “Billion-edge scale” without end-to-end workload and resource measurements is not a reproducible performance result. Mark timeouts and out-of-memory failures with the comparison budget and conditions.

Industry deployment and online experiments are useful where they support the contribution, not universal requirements. Report business or user effects only when the study design and access permissions support them. AUC or benchmark accuracy does not establish production cost savings or causal user benefit.

Drafting pattern (illustrative, unexecuted)

Problem: [task and setting]
Gap: [specific limitation, with verified prior work]
Contribution: [method / evidence / resource / design knowledge]
Evidence: [study or analysis with actual results and uncertainty]
Scope: [population, languages, workload, or conditions evaluated]
Limitations: [remaining uncertainty and unsupported settings]
Availability: [actual code/data/material access conditions]

A contribution list is optional writing structure unless the call says otherwise. See reviewer_expectations.md for preparing a response under the actual review rules, and ml_conference_style.md for quantitative ML evaluation details.