A CIKM paper is read in three accents at once. The IR reviewer scans for the task and the evaluation; the mining reviewer scans for the mechanism and its delta; the KM/database reviewer scans for the data reality — schemas, scale, heterogeneity, noise. Style at this venue means making all three scans succeed in the first page, inside a budget where appendices offer no escape (10 pages including them for 2026 full papers).
Order the opening so each lane hits its anchor early:
If the paper cannot state point 2, it is a single-lane paper and
cikm-topic-selection should re-run before more polishing.
Cross-community words are where CIKM drafts leak credibility. Terms with lane- specific meanings — entity, knowledge, relevance, scalable, online — must be pinned at first use. A table of rewrites that recur in CIKM drafts:
| Draft habit | Lane it annoys | Repair |
|---|---|---|
| "our knowledge-aware model" (never defined) | KM-DB | Name the structure: "a model conditioned on a curated product KG (2.1M triples)" |
| "significantly better" without a test | IR | Reserve the word for tested significance; else give the margin |
| "scales to large graphs" | mining | Give the size, hardware, and growth curve actually measured |
| "real-world dataset" | all three | Name the source, size, license, and collection window |
| "state-of-the-art baseline" | mining + IR | Cite venue + year; stale SOTA claims are checked at review |
Because appendices count inside the 2026 budgets, the compression question is not "what moves to the appendix" but "what leaves the PDF." Priority order for cuts: redundant pipeline diagrams; per-dataset tables that repeat one pattern (keep the pattern statement plus the fullest table); hyperparameter grids (→ the cited artifact); derivations of standard results (→ cite). What must stay: everything a reviewer needs to not trust you — evaluation protocol, baseline tuning parity, and the one ablation that isolates the mechanism.
Double-blind style at CIKM includes third-person self-citation ("Cao et al. proposed", even when Cao is an author) and de-identified system prose for applied papers — the deployment must stay describable without becoming attributable. The GenAI Usage Disclosure section is also prose: write it in the same factual register as the experiments ("evaluation scripts drafted with a code assistant, reviewed and tested by the authors; manuscript text human-written"), not as legal boilerplate.
A load-bearing default under the appendix-inclusive budget (adjust, don't worship):
1.0 pp Introduction — the four-step tri-lane opening
0.75 pp Related work — three fronts + the boundary-work paragraph
2.5 pp Method — problem formulation, mechanism, complexity note
3.0 pp Experiments — protocol ¶, main results, decisive ablation, robustness
0.5 pp Deployment/resource specifics (track-dependent)
0.5 pp Limitations + conclusion
1.75 pp Appendix (counts!) — only review-critical overflow
— GenAI Usage Disclosure + references (uncounted)
The commonest imbalance in CIKM drafts is a method section eating four-plus pages while the evaluation protocol gets one paragraph; the blended panel reads the protocol paragraph first.
Title and abstract are consumed twice before any human reads page 2: by EasyChair bidding (which lane's reviewers self-select) and later by ACM DL search. A CIKM title should carry the task, the data structure, and — when honest — the setting: "Freshness-Aware Entity Linking for Enterprise Search" routes correctly; "Towards Better Knowledge Utilization" routes randomly. In the abstract, spend one sentence per lane anchor, and put the one number you want remembered in the final sentence — that sentence is what shows in search snippets.
Draft: "With the explosion of information, knowledge graphs have attracted much attention. We propose a novel framework, XYZ, which significantly outperforms state-of-the-art methods."
Rewritten for the tri-lane scan: "Product search at catalog scale fails on queries whose intent lives in relations the index never sees — compatibility, succession, bundling. We link the query to a 2.1M-triple product KG at retrieval time and propagate relevance along typed edges; on two public benchmarks and one industrial log this lifts nDCG@10 by 4-6 points over the strongest KG-free ranker, with the gain concentrated exactly on relation-dependent queries." Problem, boundary, mechanism, calibrated evidence — one breath each.
A limitations paragraph that works at CIKM scopes rather than confesses: state the regime the evidence covers (data types, scale, languages, graph density), name the adjacent regime it does not, and say what would be needed to extend — three sentences, no adjectives of apology. The blended panel treats a well-scoped limitation as expertise and an unscoped one ("may not generalize") as either evasion or, worse, an invitation to imagine failures the authors never tested.
[Three-scan check] IR anchor / mining anchor / KM-DB anchor — found on page 1?
[Pinned terms] <cross-lane words defined at first use>
[Claim-evidence gaps] <sentences promising more than the tables hold>
[Cut list] <what leaves the PDF to make the appendix-inclusive budget>
[Compliance prose] <self-citation person, deployment de-identification, GenAI section tone>