Use this after initial ICML reviews arrive. The response should clarify and correct; it cannot rely on uploading a revised paper during the feedback period.
ICML 2026 author instructions state that authors can respond after initial reviews, that reviewers see those responses after the response period, and that reviewers and authors may then have one additional communication round. Responses must remain anonymous and should not contain non-anonymized URLs, personal websites, or shortened URLs. Reviewer instructions tell reviewers to explain whether the rebuttal changed their final recommendation.
ICML gives one response window plus one discussion round, so spend words on what an AC can quote, not on every minor note.
| Reviewer objection | Best ICML response move |
|---|---|
| "Assumptions do not match the experiments" | Concede the regime, point to the in-regime figure, state the camera-ready clarification |
| "Missing tuned baseline" | Supply a single decisive table row from a feasible run |
| "Unclear novelty vs concurrent work" | Restate the one-line technical delta and the appendix PDF |
| "Reproducibility unclear" | Cite the exact supplement command and seed |
Reviewer 2 says the convergence proof assumes bounded variance that ImageNet training violates. The strongest response opens with the concession, narrows the theoretical claim to the stated regime, adds a small plot showing the rate still tracks empirically where assumptions hold, and promises only a camera-ready wording fix. It does not argue the score or upload a revised PDF, both of which the ICML response model disallows.
[Priority issue] <reviewer/AC concern>
[Decision dimension] soundness / originality / significance / clarity / ethics / reproducibility
[Draft response] <ICML-ready anonymous text>
[Evidence anchor] <paper/appendix/supplement item>
[Discussion follow-up] <one possible reply if reviewer pushes back>