icml-review-process

ICML Review Process

Use this to reason about how ICML reviewers and ACs will process a submission. The point is not to

forecast acceptance by score folklore; it is to identify the decision issue and response strategy.

Process features

• ICML uses OpenReview with double-blind review.
• ICML 2026 aims for about three reviews per paper, though the number can vary.
• Reviewers consider soundness, originality, significance, clarity, and whether rebuttal resolved

their main concerns.

• Authors get a response period and then one additional reviewer-author discussion round.
• Reviewers and ACs may discuss privately; ACs synthesize conflicts and unresolved concerns.
• Reviews and author-reviewer discussion for accepted papers are made public.
• ICML 2026 uses a Policy A / Policy B framework for reviewer LLM use, with authors indicating

whether their paper requires Policy A review.

• Reciprocal reviewing links author behavior to reviewer obligations and policy compliance.

Diagnosis workflow

1. Categorize each likely criticism by soundness, originality, significance, clarity, ethics,

reproducibility, or policy.

1. Identify which criticism an AC can quote in a meta-review.
2. Separate fixable clarity issues from missing evidence.
3. Prepare response text that helps reviewers update their final justification.
4. Check whether LLM-review policy, prompt injection, reciprocal reviewing, or ethics concerns could

become procedural rejection risks.

Score-dimension to AC-leverage map

ICML ACs synthesize conflicts into a meta-review, so the useful question is which dimension an AC can

defend a decision on.

| Dimension | What a reviewer cites | What an AC can act on |

| --- | --- | --- |

| Soundness | Broken proof, untuned baseline, leakage | Strong reject anchor; hard to rebut without new evidence |

| Originality | Overlap with prior or concurrent work | Reject anchor unless delta is reframed |

| Significance | Narrow gain, no mechanism | Borderline; rebuttal can reframe scope |

| Clarity | Buried contribution, dense columns | Often fixable; weakest reject basis alone |

Worked vignette: reading a split panel

The optimizer paper draws one champion citing the convergence theorem, one skeptic citing an

untuned Adam baseline, and one borderline reviewer citing clarity. The AC-quotable issue is the

baseline soundness concern, not clarity, so the response strategy concentrates the single discussion

round on the tuned-baseline table. Because accepted reviews and discussion become public, the authors

keep the tone measured. If the paper also flagged Policy A LLM-review handling, confirm that choice is

consistent so it does not surface as a procedural risk. Verify the per-paper review count and policy

labels against the current cycle rather than assuming the prior year's setup.

Output format

[Likely review pattern] supportive / split / skeptical
[AC decision issue] <one issue>
[Procedural risks] <LLM policy / reciprocal reviewer / ethics / anonymity / none>
[Response posture] clarify / add small result / concede / reroute
[Public-record concern] <what will be visible if accepted>