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>
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---56# COLT Review Process78Use this to plan around COLT's review pipeline. Process facts below were verified9against the COLT 2026 CFP on 2026-07-08; mechanics are re-decided by each edition's10program chairs, so reconfirm in the cycle you are in.1112## The 2026 pipeline1314- Submission via Microsoft CMT by February 4, 2026 (AoE) for the 39th edition.15- Double-anonymous refereeing with a twist: reviewers do not see author names, but the16 area chair handling the paper does, and may reveal identities to a reviewer during17 the rebuttal period on request when needed for a proper review.18- Initial reviews go to authors before decisions; a rebuttal window follows.19- Accepted papers appear in PMLR (v291 carried COLT 2025; the 2026 volume number is20 assigned at publication).21- COLT is run by the Association for Computational Learning, with program chairs22 rotating yearly; the 2026 chair names were not published in a form verifiable at the23 access date (待核实 — check learningtheory.org/colt2026/).2425## Who reviews a COLT paper2627The pool is learning theorists: statistical learning, online learning and bandits,28optimization theory, RL theory, privacy, and adjacent CS theory. Practical29consequences:3031- Expect at least one reviewer working within epsilon of your subfield who will32 actually verify proofs, not skim them.33- "Standard techniques" claims get adjudicated by people who know the standard34 techniques' exact reach — bluffing about novelty of technique fails here faster than35 anywhere else in ML.36- Empirical framing earns nothing by itself; a reviewer may value an illustrative37 experiment, but no reviewer will accept one in lieu of a proof.3839## What the scores actually track4041| Dimension | Raises it | Sinks it |42|---|---|---|43| Correctness | Complete proofs, tracked constants, checked external results | One irreparable gap — usually terminal regardless of other strengths |44| Significance | Resolving a known open question; improving a known rate; a clean new model | A bound in a model nobody asked about, unmotivated |45| Novelty of technique | An argument that visibly cannot be assembled from known parts | A reduction the reviewer completes in the margin |46| Tightness / completeness | Matching upper and lower bounds; explicit regime coverage | Upper bound only, gap to the known lower bound unexamined |47| Clarity | Formal setup early, roadmap per proof, stable notation | Definitions scattered; appendix required to parse the theorem |4849## Decision dynamics5051- One sustained correctness objection outweighs any number of enthusiasm points; the52 AC's first job is deciding whether the proofs close.53- Because the AC knows author identities, arguments in rebuttal should stand on54 mathematics alone — appeals to seniority or track record are visible and land badly.55- Significance disputes ("who cares about this model?") are where rebuttals genuinely56 move decisions: a crisp paragraph tying the model to a named prior line, an open57 problem, or an empirical phenomenon can flip a fence-sitter.58- COLT is single-track and relatively small; acceptance implies your talk enters the59 entire community's field of view, and the bar reflects that.60- Some editions have used an "accept with minor revisions verified by the AC" flavor61 of shepherding for fixable issues; whether the current cycle does is announced with62 decisions (待核实).6364## Reading a COLT review6566```text67Review anatomy — where the decision signal lives:681. Summary paragraph -> did the reviewer parse the model correctly?69 If not, your rebuttal's first job is the misread.702. "Detailed comments" -> line-numbered proof remarks; each is a verification71 trace. Silence about App. C means C was not read.723. Questions to authors -> the actual decision hinges; answer these first.734. Typo list -> free labor; acknowledge briefly, fix silently.74```7576Reviews that engage deeply with the proofs are good news even when negative — the77paper was taken seriously, and precise objections are answerable. The dangerous review78is the short, high-level one; it signals the significance case never landed, and that79is a framing problem the rebuttal must solve.8081## After the decision: reading the outcome8283- **Accept:** the reviews still matter — camera-ready promises made in rebuttal are84 expected in the final PDF, and the AC may check (`colt-camera-ready` keeps the85 ledger).86- **Reject with proof-level reviews:** you received a free verification report from87 three experts. Patch the mathematics, then choose between the next COLT, ALT88 (the same community's sister venue, roughly anti-phased deadline), or a journal if89 the fixed version grew.90- **Reject with significance objections only:** the theorems survived; the framing91 died. Rebuild the known-vs-new ledger and the model-motivation paragraph before92 resubmitting anywhere — the same reviewers may see it again in a small community.93- **Reject you believe is wrong:** there is no formal appeals process to rely on;94 the productive channel is a stronger paper, since the community re-reviews95 resubmissions on their mathematics.9697## Confidentiality and conduct9899- Submissions are confidential; reviewers may not use or share the results before100 publication.101- COLT publishes a code of conduct for participants (a 2026 page exists at102 learningtheory.org); professional-conduct expectations extend to the rebuttal tone.103- Reviewer conflicts run through CMT domain and coauthor declarations — enter them104 completely, since the informed-AC model depends on accurate conflict data.105- Discussing your submission publicly (talks, social media) during review is not106 forbidden by anonymity rules aimed at reviewers, but volume-seeking publicity107 during the review window is poor form in a community this small.108109## Cycle-volatility warnings110111- Portal, anonymity mechanics, rebuttal format, shepherding, and decision timeline are112 annual decisions. Decision dates for 2026 were not on the pages checked (待核实).113- Reviewer-volunteering expectations for submitting authors have not been a stated114 COLT policy in the verified material; do not assume one either way without the115 current CFP.116117## Output format118119```text120[Stage] pre-submission / under review / rebuttal / decision121[Score driver] correctness / significance / technique / tightness / clarity122[Reviewer engagement] deep proof-level / shallow-summary (framing failed)123[Rebuttal leverage] <the one thread that can move the decision>124[Process facts to reconfirm] <current-cycle items still 待核实>125```
brycewang-stanford