brycewang-stanford ATC Author Response
ATC gives you two distinct writing moments after review, and they have different jobs. The rebuttal (a short, roughly one-page window during round two) is for correcting misreadings and supplying a measurement a reviewer said was missing. The description of changes (after a conditional accept, for the shepherd) is a point-by-point record proving every required revision was made. Confusing the two — arguing in the shepherd round, or over-promising in the rebuttal — is the classic failure.
The rebuttal (round two, ~1 page)
Its job is narrow and its space is tiny, so triage ruthlessly:
- Correct factual misreadings first — a reviewer who misread the testbed, the baseline, or the claim can be moved by a precise correction with a pointer to the section/figure.
- Supply the one number that matters — if a reviewer said "no tail latency" or "no matched-cost comparison," and you have or can add it, state the result concisely. Systems reviewers move on evidence, not on rhetoric.
- Concede cleanly what you cannot fix this cycle — a crisp "correct; here is the honest bound" buys more credibility than defending an overclaim.
- Do not argue taste — "the reviewer underrates the contribution" changes nothing; the PC discussion decides.
Structure it as reviewer/point → response, lead with the highest-impact corrections, and never introduce unsupported new claims (which read as scope creep). Keep double-blind: the rebuttal must not reveal identity.
The description of changes (shepherd round)
After a conditional accept, a PC member shepherds the paper: final acceptance depends on the required revisions landing. Write the change record as a contract:
- One row per required change: reviewer/shepherd request → what you changed → where (section, figure, table) → the resulting text or result.
- Make or explicitly decline every requested change. A silently skipped required change is the fastest way to stall a conditional accept; if you decline, give the reason and the evidence.
- Point to concrete artifacts of the change — the new figure, the added experiment, the revised paragraph — not a promise to do it.
- Keep the shepherd's scope. The shepherd enforces the required revisions, not a new round of wishes; do not reopen settled scope or add unrequested claims that invite fresh review.
Mapping requests to moves
| Reviewer request | Strong response |
|---|---|
| "Baseline is weak / unmatched" | Add the well-configured baseline at a matched operating point; report the new comparison |
| "No tail latency / variance" | Add p99/p999 and multi-run variance; point to the updated figure |
| "Unclear how it's built / scales" | Add implementation detail or a scaling experiment; cite the new section |
| "Costs not reported" | Quantify the trade beside the gain at matched cost |
| "Reproducibility unclear" | Point to the anonymized artifact and the claim-to-experiment map |
| A misreading | Correct precisely with a section/figure pointer; do not over-explain |
What moves an outcome vs. what does not
[Moves it] a factual correction; a requested measurement supplied; a clean concession + bound
[Doesn't] arguing the reviewer's taste; re-asserting the abstract; vague "we will consider"
[Backfires] unsupported new claims; scope creep; defensiveness; an anonymity leak
Reverify each cycle
- Whether a formal rebuttal window exists and its length (about one page is the reported norm — confirm on the call).
- The shepherding requirements and the change-record format for the current edition.
- The camera-ready deadline relative to the shepherd sign-off (see
atc-camera-ready).
Output format
[Response type] round-two rebuttal / shepherd description-of-changes
[Triage] highest-impact corrections and requested measurements listed first
[Per point] request -> response -> evidence/location; every required change made or declined-with-reason
[Anonymity] no identity leak in the response? yes/no
[Forbidden] unsupported new claims / taste arguments / silently skipped required changes: absent?