COLT Review Process
Use this to plan around COLT's review pipeline. Process facts below were verified
against the COLT 2026 CFP on 2026-07-08; mechanics are re-decided by each edition's
program chairs, so reconfirm in the cycle you are in.
The 2026 pipeline
- Submission via Microsoft CMT by February 4, 2026 (AoE) for the 39th edition.
- Double-anonymous refereeing with a twist: reviewers do not see author names, but the
area chair handling the paper does, and may reveal identities to a reviewer during
the rebuttal period on request when needed for a proper review.
- Initial reviews go to authors before decisions; a rebuttal window follows.
- Accepted papers appear in PMLR (v291 carried COLT 2025; the 2026 volume number is
assigned at publication).
- COLT is run by the Association for Computational Learning, with program chairs
rotating yearly; the 2026 chair names were not published in a form verifiable at the
access date (待核实 — check learningtheory.org/colt2026/).
Who reviews a COLT paper
The pool is learning theorists: statistical learning, online learning and bandits,
optimization theory, RL theory, privacy, and adjacent CS theory. Practical
consequences:
- Expect at least one reviewer working within epsilon of your subfield who will
actually verify proofs, not skim them.
- "Standard techniques" claims get adjudicated by people who know the standard
techniques' exact reach — bluffing about novelty of technique fails here faster than
anywhere else in ML.
- Empirical framing earns nothing by itself; a reviewer may value an illustrative
experiment, but no reviewer will accept one in lieu of a proof.
What the scores actually track
| Dimension |
Raises it |
Sinks it |
| Correctness |
Complete proofs, tracked constants, checked external results |
One irreparable gap — usually terminal regardless of other strengths |
| Significance |
Resolving a known open question; improving a known rate; a clean new model |
A bound in a model nobody asked about, unmotivated |
| Novelty of technique |
An argument that visibly cannot be assembled from known parts |
A reduction the reviewer completes in the margin |
| Tightness / completeness |
Matching upper and lower bounds; explicit regime coverage |
Upper bound only, gap to the known lower bound unexamined |
| Clarity |
Formal setup early, roadmap per proof, stable notation |
Definitions scattered; appendix required to parse the theorem |
Decision dynamics
- One sustained correctness objection outweighs any number of enthusiasm points; the
AC's first job is deciding whether the proofs close.
- Because the AC knows author identities, arguments in rebuttal should stand on
mathematics alone — appeals to seniority or track record are visible and land badly.
- Significance disputes ("who cares about this model?") are where rebuttals genuinely
move decisions: a crisp paragraph tying the model to a named prior line, an open
problem, or an empirical phenomenon can flip a fence-sitter.
- COLT is single-track and relatively small; acceptance implies your talk enters the
entire community's field of view, and the bar reflects that.
- Some editions have used an "accept with minor revisions verified by the AC" flavor
of shepherding for fixable issues; whether the current cycle does is announced with
decisions (待核实).
Reading a COLT review
Review anatomy — where the decision signal lives:
1. Summary paragraph -> did the reviewer parse the model correctly?
If not, your rebuttal's first job is the misread.
2. "Detailed comments" -> line-numbered proof remarks; each is a verification
trace. Silence about App. C means C was not read.
3. Questions to authors -> the actual decision hinges; answer these first.
4. Typo list -> free labor; acknowledge briefly, fix silently.
Reviews that engage deeply with the proofs are good news even when negative — the
paper was taken seriously, and precise objections are answerable. The dangerous review
is the short, high-level one; it signals the significance case never landed, and that
is a framing problem the rebuttal must solve.
After the decision: reading the outcome
- Accept: the reviews still matter — camera-ready promises made in rebuttal are
expected in the final PDF, and the AC may check (
colt-camera-ready keeps the
ledger).
- Reject with proof-level reviews: you received a free verification report from
three experts. Patch the mathematics, then choose between the next COLT, ALT
(the same community's sister venue, roughly anti-phased deadline), or a journal if
the fixed version grew.
- Reject with significance objections only: the theorems survived; the framing
died. Rebuild the known-vs-new ledger and the model-motivation paragraph before
resubmitting anywhere — the same reviewers may see it again in a small community.
- Reject you believe is wrong: there is no formal appeals process to rely on;
the productive channel is a stronger paper, since the community re-reviews
resubmissions on their mathematics.
Confidentiality and conduct
- Submissions are confidential; reviewers may not use or share the results before
publication.
- COLT publishes a code of conduct for participants (a 2026 page exists at
learningtheory.org); professional-conduct expectations extend to the rebuttal tone.
- Reviewer conflicts run through CMT domain and coauthor declarations — enter them
completely, since the informed-AC model depends on accurate conflict data.
- Discussing your submission publicly (talks, social media) during review is not
forbidden by anonymity rules aimed at reviewers, but volume-seeking publicity
during the review window is poor form in a community this small.
Cycle-volatility warnings
- Portal, anonymity mechanics, rebuttal format, shepherding, and decision timeline are
annual decisions. Decision dates for 2026 were not on the pages checked (待核实).
- Reviewer-volunteering expectations for submitting authors have not been a stated
COLT policy in the verified material; do not assume one either way without the
current CFP.
Output format
[Stage] pre-submission / under review / rebuttal / decision
[Score driver] correctness / significance / technique / tightness / clarity
[Reviewer engagement] deep proof-level / shallow-summary (framing failed)
[Rebuttal leverage] <the one thread that can move the decision>
[Process facts to reconfirm] <current-cycle items still 待核实>
Source: brycewang-stanford/Awesome-Journal-Skills → COLT-Skills/skills/colt-review-process/SKILL.md
1---2name: colt-review-process3description: Use when reasoning about COLT (Conference on Learning Theory) peer review — the double-anonymous-with-informed-area-chair model, correctness-first evaluation by expert theorists, the rebuttal stage before decisions, single-track acceptance stakes, and how PMLR publication and the community's proof culture shape outcomes.4---567# COLT Review Process89Use this to plan around COLT's review pipeline. Process facts below were verified10against the COLT 2026 CFP on 2026-07-08; mechanics are re-decided by each edition's11program chairs, so reconfirm in the cycle you are in.1213## The 2026 pipeline1415- Submission via Microsoft CMT by February 4, 2026 (AoE) for the 39th edition.16- Double-anonymous refereeing with a twist: reviewers do not see author names, but the17 area chair handling the paper does, and may reveal identities to a reviewer during18 the rebuttal period on request when needed for a proper review.19- Initial reviews go to authors before decisions; a rebuttal window follows.20- Accepted papers appear in PMLR (v291 carried COLT 2025; the 2026 volume number is21 assigned at publication).22- COLT is run by the Association for Computational Learning, with program chairs23 rotating yearly; the 2026 chair names were not published in a form verifiable at the24 access date (待核实 — check learningtheory.org/colt2026/).2526## Who reviews a COLT paper2728The pool is learning theorists: statistical learning, online learning and bandits,29optimization theory, RL theory, privacy, and adjacent CS theory. Practical30consequences:3132- Expect at least one reviewer working within epsilon of your subfield who will33 actually verify proofs, not skim them.34- "Standard techniques" claims get adjudicated by people who know the standard35 techniques' exact reach — bluffing about novelty of technique fails here faster than36 anywhere else in ML.37- Empirical framing earns nothing by itself; a reviewer may value an illustrative38 experiment, but no reviewer will accept one in lieu of a proof.3940## What the scores actually track4142| Dimension | Raises it | Sinks it |43|---|---|---|44| Correctness | Complete proofs, tracked constants, checked external results | One irreparable gap — usually terminal regardless of other strengths |45| Significance | Resolving a known open question; improving a known rate; a clean new model | A bound in a model nobody asked about, unmotivated |46| Novelty of technique | An argument that visibly cannot be assembled from known parts | A reduction the reviewer completes in the margin |47| Tightness / completeness | Matching upper and lower bounds; explicit regime coverage | Upper bound only, gap to the known lower bound unexamined |48| Clarity | Formal setup early, roadmap per proof, stable notation | Definitions scattered; appendix required to parse the theorem |4950## Decision dynamics5152- One sustained correctness objection outweighs any number of enthusiasm points; the53 AC's first job is deciding whether the proofs close.54- Because the AC knows author identities, arguments in rebuttal should stand on55 mathematics alone — appeals to seniority or track record are visible and land badly.56- Significance disputes ("who cares about this model?") are where rebuttals genuinely57 move decisions: a crisp paragraph tying the model to a named prior line, an open58 problem, or an empirical phenomenon can flip a fence-sitter.59- COLT is single-track and relatively small; acceptance implies your talk enters the60 entire community's field of view, and the bar reflects that.61- Some editions have used an "accept with minor revisions verified by the AC" flavor62 of shepherding for fixable issues; whether the current cycle does is announced with63 decisions (待核实).6465## Reading a COLT review6667```text68Review anatomy — where the decision signal lives:691. Summary paragraph -> did the reviewer parse the model correctly?70 If not, your rebuttal's first job is the misread.712. "Detailed comments" -> line-numbered proof remarks; each is a verification72 trace. Silence about App. C means C was not read.733. Questions to authors -> the actual decision hinges; answer these first.744. Typo list -> free labor; acknowledge briefly, fix silently.75```7677Reviews that engage deeply with the proofs are good news even when negative — the78paper was taken seriously, and precise objections are answerable. The dangerous review79is the short, high-level one; it signals the significance case never landed, and that80is a framing problem the rebuttal must solve.8182## After the decision: reading the outcome8384- **Accept:** the reviews still matter — camera-ready promises made in rebuttal are85 expected in the final PDF, and the AC may check (`colt-camera-ready` keeps the86 ledger).87- **Reject with proof-level reviews:** you received a free verification report from88 three experts. Patch the mathematics, then choose between the next COLT, ALT89 (the same community's sister venue, roughly anti-phased deadline), or a journal if90 the fixed version grew.91- **Reject with significance objections only:** the theorems survived; the framing92 died. Rebuild the known-vs-new ledger and the model-motivation paragraph before93 resubmitting anywhere — the same reviewers may see it again in a small community.94- **Reject you believe is wrong:** there is no formal appeals process to rely on;95 the productive channel is a stronger paper, since the community re-reviews96 resubmissions on their mathematics.9798## Confidentiality and conduct99100- Submissions are confidential; reviewers may not use or share the results before101 publication.102- COLT publishes a code of conduct for participants (a 2026 page exists at103 learningtheory.org); professional-conduct expectations extend to the rebuttal tone.104- Reviewer conflicts run through CMT domain and coauthor declarations — enter them105 completely, since the informed-AC model depends on accurate conflict data.106- Discussing your submission publicly (talks, social media) during review is not107 forbidden by anonymity rules aimed at reviewers, but volume-seeking publicity108 during the review window is poor form in a community this small.109110## Cycle-volatility warnings111112- Portal, anonymity mechanics, rebuttal format, shepherding, and decision timeline are113 annual decisions. Decision dates for 2026 were not on the pages checked (待核实).114- Reviewer-volunteering expectations for submitting authors have not been a stated115 COLT policy in the verified material; do not assume one either way without the116 current CFP.117118## Output format119120```text121[Stage] pre-submission / under review / rebuttal / decision122[Score driver] correctness / significance / technique / tightness / clarity123[Reviewer engagement] deep proof-level / shallow-summary (framing failed)124[Rebuttal leverage] <the one thread that can move the decision>125[Process facts to reconfirm] <current-cycle items still 待核实>126```127128---129130**Source:** [`brycewang-stanford/Awesome-Journal-Skills`](https://github.com/brycewang-stanford/Awesome-Journal-Skills) → `COLT-Skills/skills/colt-review-process/SKILL.md`
thedixitjain