kdd-author-response
GitHub用于起草KDD论文Rebuttal，遵循无超链接规则。涵盖证据定位、按 reviewer 分类处理（事实纠正、基线对比等）、针对AC的决策关键回应，以及将Resubmit转化为下一周期材料的策略。
Trigger Scenarios
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill kdd-author-response -g -y
SKILL.md
Frontmatter
{
"name": "kdd-author-response",
"description": "Use when drafting KDD rebuttals on OpenReview, where authors answer each review under a no-hyperlink rule, cannot upload revisions, and write for the area chair who recommends Accept, Reject, or Resubmit. Covers evidence-anchoring without links, per-reviewer triage, and turning a likely Resubmit into next-cycle material at ACM SIGKDD."
}
KDD Author Response
Use this when KDD reviews land. The structural facts that shape every rebuttal here: authors get a chance to respond to each review; the area chair reads the responses plus the reviewer discussion and recommends; the PC chairs decide; and the decision space includes Resubmit, not just accept/reject. Confirm the current cycle's exact rebuttal window and length limits on OpenReview before drafting — mechanics move between cycles.
The no-links constraint
KDD 2026 forbade hyperlinks in rebuttals and discussion. Practical consequences:
- You cannot introduce a new repository, demo, or anonymous site now. The only artifact reviewers can check is whatever the submitted PDF already references.
- Point with coordinates instead: "Table 3, column 4", "Section 5.2, lines 412-419"
(the
reviewoption's line numbers exist precisely for this), "Appendix B, Algorithm 2", "README section 'Repro' of the repository cited in Section 6". - New numbers may be typed into the response as plain text tables if the cycle's rules allow new experimental results at all — check before running anything.
Per-reviewer triage
Sort every reviewer point into one of four bins before writing a word:
| Bin | Example at KDD | Response move |
|---|---|---|
| Factual misread | "No ablation for the drift module" when Table 5 is that ablation | Correct politely with exact coordinates; this is the cheapest score repair available |
| Missing-baseline demand | "Compare against the latest graph transformer" | Run it if rules and time permit and report as text; otherwise explain the regime mismatch concretely |
| Scale skepticism | "Would this hold beyond the 10M-edge graphs tested?" | Cite the complexity analysis plus the largest reported run; concede the untested regime explicitly |
| Scope objection | "This is engineering, not research" (or, on ADS: "no deployment evidence") | Restate the contribution as mechanism + regime; for ADS point to the post-launch metrics section |
Answer the area chair, not the loudest reviewer: one resolved decision-critical objection moves a recommendation more than five resolved nitpicks.
Drafting pattern per review
R<k>, thank you for the careful read. We respond to your main concerns:
[W1 - <reviewer's phrase>] <one-sentence answer>.
Evidence: Section X.Y / Table N / Appendix Z (submitted version).
<If conceded:> This is correct; we will scope the claim in the final version to <...>.
[W2 - ...] ...
Summary: the decision-relevant points are W1 (misread, resolved by Table N) and
W3 (new run reported above as plain text). We believe W2 is a presentation fix
we commit to for the final version.
Keep each response self-contained; area chairs skim per-review threads independently.
Playing for Resubmit when Accept is out of reach
Unique to KDD's dual-cycle system: a rebuttal that will not rescue acceptance can still set up a strong next-cycle resubmission, because the Resubmit path expects the noted concerns to be addressed and rewards papers that address them.
- Concede precisely. A concession recorded in this cycle's thread becomes the first line of the next cycle's mandatory one-page change summary.
- Ask for actionability: if a reviewer's objection is vague, a polite request to name the missing experiment gives the resubmission a concrete target.
- Never burn the thread. The resubmission declares the previous forum id, so hostile rebuttal text travels with the paper into the next cycle.
Tone calibration for this venue
- KDD reviewer pools mix academic miners with industry practitioners; answer efficiency and deployment questions with numbers, not with "left to future work".
- Do not promise camera-ready experiments you have not run — ACs discount unbacked promises, and the Resubmit option means "come back when it exists" is a real outcome.
- Confidentiality and anonymity still bind: no identity hints, no "as our production
system at
" phrasing in Research Track responses.
Worked micro-exchange
Reviewer 2 (practitioner lens) writes: "The 12% lift over the baseline likely comes from the extra embedding table, not the proposed propagation rule. Also, results on the industrial dataset cannot be verified." A response that moves the AC:
- Attribute with the ablation: "The variant
-propagationin Table 5 keeps the embedding table and removes only the propagation rule; the lift drops from 12% to 3%, isolating the rule's contribution (Section 5.3, lines 512-528)." - Convert the verification complaint into a specification answer: "The industrial dataset cannot be released, but Section 4.1 specifies schema, size (1.9B interactions), and collection window, and all pipeline code runs on the public datasets in the repository cited in Section 6."
- Concede the residue: "We agree the 3% residual on public data is smaller than the industrial figure and will state this asymmetry explicitly in the final version."
Three moves, three sentences each, zero links — and each maps a review objection to submitted evidence or a bounded commitment.
Rebuttal-window logistics
- Draft against the clock: per-review responses for a 3-4 reviewer packet under a
length limit take a full author-day; two for papers with new runs. Reserve the
capacity when reviews are announced, not when they arrive (
kdd-workflow). - One author owns tone-consistency across all responses; reviewers compare notes in discussion, and contradictory answers to R1 and R3 are visible to the AC.
- Character/word limits per response box vary by cycle (待核实) — measure before writing, since cutting a finished response degrades it more than planning short.
- After submitting responses, monitor the forum if the cycle allows follow-up discussion; an unanswered reviewer follow-up in the closing days reads as abandonment.
Output format
[Rebuttal stance] repair-for-accept / stabilize-borderline / build-for-resubmit
[Per-reviewer bins] R1: <bin>, R2: <bin>, R3: <bin>
[Decision-critical point] <the one objection the AC will weigh>
[Evidence anchors] <section/table/line coordinates only - no links>
[Concessions recorded] <items that feed a Cycle-2 change summary>
[Forbidden content check] links: none / identity: clean / unbacked promises: none
Version History
- 9f86f09 Current 2026-07-19 16:56
