colt-author-response
GitHub用于COLT论文初评后撰写反驳信。指导如何修复证明漏洞、应对紧致性与新颖性质疑，遵守AC身份揭示规则，避免引入未验证新定理，并提供针对理论审稿人常见异议的标准应对策略与写作范式。
Trigger Scenarios
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill colt-author-response -g -y
SKILL.md
Frontmatter
{
"name": "colt-author-response",
"description": "Use when drafting a COLT (Conference on Learning Theory) rebuttal after initial reviews, handling proof-correctness objections, tightness and novelty-of-technique challenges, the area-chair identity-reveal rule, and concession strategy for theorem-first papers where one unrepaired gap outweighs every polished answer."
}
COLT Author Response
Use this after COLT reviews arrive. The COLT 2026 CFP (checked 2026-07-08) stated that initial reviews are sent to authors before final decisions and that authors get an opportunity to address issues raised — a rebuttal phase, run through CMT in that cycle. Exact window length, character limits, and attachment rules are announced with the reviews; read that announcement before writing a word.
What a COLT rebuttal is for
A COLT review stands or falls on correctness first, then on the significance of the result and the novelty of the technique. The rebuttal is your one chance to:
- repair a perceived gap by pointing at the exact appendix line that closes it;
- concede a real gap and show the damage is contained (a weaker constant, an extra log factor, a mildly stronger assumption) with a sketch of the fix;
- correct a factual misreading of the model, the adversary, or the oracle access;
- defuse "this follows easily from known results" by naming precisely which step known techniques cannot deliver.
It is not a place to announce new theorems. A brand-new result in a rebuttal box is unverifiable in the time reviewers have, and theory reviewers discount it accordingly; promise it for the final version only when the argument is a routine extension.
The identity-reveal rule (distinctively COLT)
Under the 2026 policy, submissions are anonymized and names are withheld from reviewers, but the area chair knows author identities and may reveal them to a reviewer during the rebuttal period upon the reviewer's request, when the AC judges it necessary for a proper review — typically to check adjacency to the authors' own prior work. Consequences for drafting:
- Never write anything inconsistent with your identity becoming known mid-discussion.
- Do not play games with "the authors of [7]" phrasing that collapses if unmasked.
- You still must not de-anonymize yourself proactively; the reveal is the AC's call.
Objection playbook for theorist reviewers
| Objection | What it usually means | COLT-ready move |
|---|---|---|
| "The proof of Lemma 3 has a gap at (12)" | The reviewer traced the chain and stopped | Give the missing two-line argument verbatim, or concede and show the repaired constant |
| "The bound is not tight; the lower bound is missing" | Result may be a partial story | Point to the matching lower bound if it exists; otherwise scope the claim and cite the open gap honestly |
| "This follows from [X] by a standard reduction" | Novelty-of-technique attack | Exhibit the concrete step where the reduction fails — the exact quantity [X] cannot control |
| "The assumption of bounded losses is restrictive" | Generality probe | Say whether the proof survives sub-Gaussian relaxation, and where it breaks if not |
| "The model is non-standard" | Fit and significance doubt | Anchor the model to a named line of prior COLT/ALT work and state what changes and why |
Drafting pattern
- One-line verdict per major objection: correct as stated, misreading, or conceded with contained damage — reviewers scan for whether you dodged.
- Quote the exact display or lemma number; theory reviewers verify pointers, and a wrong pointer costs more credibility than a concession.
- Keep arithmetic in the rebuttal self-contained: if the fix is three inequalities, write the three inequalities rather than "this is easy to fix."
- End each thread with what the final version will say, phrased as revision of existing material, not as new claims.
Worked micro-example
Reviewer: "Theorem 2's regret bound hides the dependence on the action-set diameter D; for unbounded D the claim is vacuous."
Verdict: misreading, with a wording fix we will make.
The D-dependence is present but implicit: the radius enters through the
self-bounding term in Eq. (9), and Lemma 5 (App. B.2) gives
R_T <= 4 D sqrt(T log K) + D log T.
D is assumed finite in Section 2 (line 3 of the setup). We agree the theorem
statement should carry D explicitly; the final version will state
R_T = O(D sqrt(T log K)) in Theorem 2 itself rather than in the proof.
Four sentences, one display, an exact pointer, one concrete promised edit — the whole anatomy of a COLT rebuttal thread.
Rebuttal anti-patterns theorists punish
- The pointer dodge: "see Appendix C" with no line number and no restatement of the resolving argument. If the referee misread C once, a bare pointer re-serves the same meal.
- The generality retreat: answering a tightness question by claiming the result "still holds in more general settings" — orthogonal to the objection and visibly so.
- The authority appeal: "this technique is standard in the literature" without a citation containing the actual step; worse if the identity-reveal rule later shows you were citing yourself.
- The scope creep: using the rebuttal to advertise results proved after submission; at best ignored, at worst it signals the submitted version was premature.
- The wounded tone: any sentence about the review's fairness rather than its mathematics. The AC reads tone as information about how camera-ready fixes would go.
- The uniform-length trap: giving every objection equal space, which tells the AC you could not identify the decision-critical one.
Prioritization under a short window
- Rank objections by decision weight: correctness gap > tightness/lower-bound gap > novelty-of-technique > model fit > exposition. Answer in that order and cut from the bottom when space runs out.
- If two reviewers disagree (one says the technique is standard, the other says the model is exotic), answer both in one paragraph that names the technique's new step — the AC reads the rebuttal for exactly this synthesis.
- A calm, complete concession of a minor error buys credibility for your defense of the main theorem; theorists reward authors who behave like referees of their own work.
Cycle-volatility warnings
- Rebuttal length limits, whether PDF responses are allowed, and per-reviewer versus global response structure change by cycle — read the current instructions (待核实 each year).
- The identity-reveal mechanism described above is the 2026 wording; verify it in the cycle you are in before relying on its exact scope.
Output format
[Priority objection] <the one that decides the paper>
[Verdict] correct as stated / misreading / conceded, damage contained
[Pointer] <lemma, display, or appendix line>
[Self-contained fix] <inequalities or sketch given in-box>
[Promised final-version edit] <existing-material revision only>
Version History
- 9f86f09 Current 2026-07-19 14:40
