Guide
COLM explained: 29% accepted, and its own chairs say it counts less
Its programme chairs wrote that COLM papers count less than other venues' because it is unranked by CORE, then that the quality is at least as high. That trade is the venue, and you take it in March.
Published 14 August 2026
Deciding whether COLM is the room? Paste the abstract lower on this page and see which venues publish work like yours. Jump to the matcher
The short version
- COLM is a single-track, archival, in-person conference on language modeling. Now in its third edition.
- It is the only venue we know of whose rubric names the compute gap. Reviewers are told to weigh that most researchers cannot buy frontier-lab compute.
- You pay for that three ways. It is unranked. Which means it reads thinner on some CVs.
- It is exclusive. An ARR paper at any stage, an ACL commitment or a live ICML submission each block you on 31 March.
- And one author has to be there in October. There is no virtual option.
Standing
Three editions, and the numbers live almost nowhere
Nearly triple the accepted papers in two years, on a tightening rate. Neither figure is on its own site.
| Figure | 2024 | 2025 | 2026 |
|---|---|---|---|
| Accepted papers | 299 | 418 | 856 |
| Submissions | Not published | 1,305 | Not published. 856 at 29% implies roughly 3,000 |
| Acceptance rate | Not published | 32% | 29% |
| Spotlights | 18 badged | 24, about 1.8% | Not published |
| Outstanding papers | 4 | 4, about 0.3% | Not published |
| Programme committee | Not published | 2,615 reviewers, 227 area chairs | Not published |
| Authors and attendance | Not published | 2,243 authors from 43 countries, about 1,400 attendees | Not published |
| Host city | Philadelphia | Montreal | San Francisco |
Accepted papers counted by us on COLM's lists for 2024, 2025 and 2026. The rest of the 2025 column is the chairs' opening-remarks deck, posted by the official COLM account on Bluesky on 7 October 2025; its 418 matches the list. The 29% is theirs, from their post on AI-written submissions. Read 14 August 2026; implied 2026 submissions are our arithmetic.
The chairs published the case against their own venue
Their wording: COLM is not ranked on rankings like CORE, so papers in COLM count less than those in other venues for some researchers, and as a result it may be less of an attractive target for gaming. They then say they believe the quality is at least as high. The ranking gap is the filter, not the flaw.
The 2025 numbers sit on one slide, and the popular source has them backwards
The aggregator that ranks for colm acceptance rate shows COLM 2025 as 418 submissions, 418 accepted, 100%, which is what scraping the accepted-paper list and calling it the submission pool gets you. The real pair is 1,305 and 418, from a deck cited nowhere else we could find. Quote 32% for 2025, 29% for 2026.
The calendar
A fork you take in March, not a fallback in June
The full-paper deadline is 31 March and it is exclusive. What you hold at ARR or ICML that day decides whether you may submit.
| Stage | COLM 2026 |
|---|---|
| Abstract deadline | Thursday 26 March 2026 |
| Full paper deadline | Tuesday 31 March 2026 |
| Author and reviewer discussion | 22 May to 8 June 2026, an 18-day window, not a fixed rebuttal slot |
| Decisions | 8 July 2026 |
| Conference | 6 to 9 October 2026, Hilton Union Square, San Francisco |
| Workshops | 9 October 2026, 18 listed |
All deadlines 23:59 AoE, from the COLM 2026 dates page, read August 2026. Across three editions: abstracts 22 to 26 March, full papers 28 to 31 March, decisions 7 to 9 July, conference in early October.
| What you hold on 31 March | COLM's rule |
|---|---|
| Under review at ICML | You must withdraw by 31 March in order to submit to COLM |
| Committed to an ACL venue | Commitment counts as under review there, so you cannot submit |
| Anywhere in the ARR process | Papers under ARR, at any stage, count as under submission there and cannot be submitted |
| Under review at another archival venue | Not permitted. Parallel submission to non-archival venues and workshops is fine |
| Posted on arXiv | Explicitly allowed, on a preprint policy adapted from NeurIPS |
Exclusivity rule from the COLM submission instructions, the ARR clause from the FAQ, read August 2026.
COLM has already moved its own deadlines over this collision
The official account announced the 2025 delay in these words: a bit of a mess around the conflict of COLM with the ARR, and to lesser degree ICML, reviews release. Abstracts went back 48 hours, full papers 24. The clash is structural: the reviews you are waiting on land in the fortnight of the deadline that forbids them. The March question is not which two venues you submit to, it is which one you are in.
How you are judged
COLM publishes a rubric, and no score sheet
Five dimensions in prose, a note that few papers if any excel along all of them, and no published form or scale anywhere.
| Dimension | What work that excels along it looks like |
|---|---|
| Empiricism, data and evaluation | Data as natural as possible, strong experimental design, metrics known or shown to measure what they claim to |
| Technological impact | High quality, thoughtfully designed, well packaged resources and artifacts that enable future work |
| Ambition, vision, forward outlook | Looks beyond the problems currently studied, or extends them in significant ways |
| Understanding depth, principled approach | Deepens understanding, by a principled approach to modeling and learning, or through deep analysis |
| Clarity, honesty and trust | Written clearly, measured and balanced, releasing research materials as far as possible |
Condensed from the COLM review guide, read August 2026, the entire published rubric. It tells reviewers to avoid mechanistic measures of quality, that weakness in some aspects cannot be an immediate reason to reject, and the inverse: vision is no panacea if the rest is weak. The only public score number is a distribution the official account posted on 24 May 2026, 50th percentile 5 and 75th percentile 5.67, on a scale never published.
No score sheet, and one instruction no other venue writes down
Two percentiles fix neither endpoint nor anchor, so anyone showing you a COLM rating scale reconstructed it. We will not: your reviewer is asked for a judgement in prose, not a threshold. What the guide says instead is sharper. Reviewers must take into account that most researchers cannot get large-scale compute, and accept the consequence: some small-scale results will not hold at scale, but some will, and will not make it unless the committee bets on them.
The fine print
The rules you cannot argue with after the deadline
Most are ordinary. The tinted one is why a COLM rejection costs less than an ICLR one.
| Rule | What COLM says |
|---|---|
| If you are rejected | Nothing is published. Rejected and withdrawn papers, their discussions and metadata stay private. Accepted papers' reviews go public after decisions |
| Length | 9 pages of main text, strict. Unlimited citations, appendices reviewers need not read, one extra page at camera-ready |
| Free sections | Ethics, reproducibility and acknowledgments, a page each, outside the limit |
| Author list | Frozen at the deadline, no modification afterwards for any reason, camera-ready included |
| Author cap | 25 submissions per person |
| Reciprocal reviewing | Every submission contributes a reviewer, no two may share one, and 4 or more submissions puts you in the pool |
| Who reviews you | Second-year graduate student experience, two peer-reviewed papers as primary author, 5 papers each |
| Who decides | Area chairs. No bidding, papers assigned by algorithm, and an AC may ignore a clearly low-quality review |
| Ethics | Ethics reviewers cannot reject, chairs may, whatever the science. In 2025: 72 flagged, 43 reviewed, 1 rejected |
| Tracks | None, no short-paper or position track, surveys in the main pool. Every paper is a poster, a few also orals |
| Attendance | One author must register. In person only, no virtual option |
Length, reviews privacy, the author cap, ethics and reciprocal reviewing from the call for papers; tracks, attendance, the reviewer bar and the author list from the FAQ; decisions from the area chair guidelines, adapted, they say, from the NeurIPS 2024 ones. Ethics counts from the chairs' deck. Read August 2026.
Enforcement
COLM 2026 desk-rejected about 2% of submissions before review
The chairs published what their AI-detection programme did. The number that matters is fourth from the bottom.
| Measure | COLM 2026 |
|---|---|
| Primarily AI-generated or very heavily AI-assisted | Roughly 5%, from a run coordinated with GPTZero |
| Papers the chairs then read by hand | Around 90, the highest-scored |
| Desk-rejected for LLM policy violations | 32 of those 90, plus 14 others outside the set |
| Desk-rejected for very poor quality or obvious hallucinations | 11 |
| Desk-rejected for completely fabricated references | Around 20 |
| Total desk-rejected on these criteria | Roughly 2% of submissions |
| Recommended for acceptance, the 50 most AI-scored papers | 0 |
| Recommended for acceptance, papers 51 to 100 | 5, which is 10%, against a 29% baseline |
| Theoryslop, a failure mode they named | A construct, a theorem or two, validation by fitting coefficients, no strong empirical core. COLM is not a theory venue |
| Slopterpretability, the other one | A narrow question, an off-the-shelf circuit or SAE analysis, conclusions sometimes exaggerated |
Figures from the COLM 2026 chairs' post on AI-written submissions, read August 2026; COLM's organizers page lists its author as senior programme chair. The call for papers adopts the ICLR 2026 LLM policy minus disclosure for minor writing or coding help; originating ideas, writing content including references, generating data or plots and LLM evaluation must be disclosed. The submission instructions add that hallucinated content in one part of a paper casts doubt on the entire paper.
No paper was rejected because a detector said so
The chairs state it twice: no paper was desk-rejected on the basis of a detector result, and no detector results were shown to reviewers or area chairs. The detector picked a reading list for humans, who rejected papers for what they found. A dataset paper with AI-assisted prose was accepted, because the usefulness of the artifact outweighed subpar, AI-trope-ridden writing.
Questions
COLM, asked directly
The questions searchers actually type, answered from the sources above and nothing else.
What is COLM's acceptance rate?
29% at COLM 2026, stated by the programme chairs, with 856 papers accepted. COLM 2025 accepted 418 of 1,305 submissions, about 32%. For 2024 no submission count was published, so no rate exists to quote. The 100% figure some aggregators show for 2025 comes from mistaking the accepted-paper list for the submission pool.
Is COLM a good conference?
By growth and selectivity, yes: accepted papers nearly tripled in two years, from 299 to 856, while the rate tightened to 29%. The honest caveat comes from its own chairs: COLM is unranked, so a COLM paper counts less on rankings-driven CVs, and they argue the quality is at least as high. Whether that trade is good depends on whether your next reader is a hiring committee or a language-modeling researcher.
Does COLM have a CORE ranking?
No. COLM's own programme chairs write that it is not ranked on rankings like CORE, and that papers here therefore count less than those at other venues for some researchers. They add that they believe the quality is at least as high. If a CORE-ranked venue is a hard requirement, that answer decides for you.
Does COLM have an impact factor?
No. Impact factor is a journal metric from Clarivate's citation index, and COLM is a conference; NeurIPS, ICML and ICLR have none either. The nearest published selectivity figure is the acceptance rate, 29% in 2026, from the programme chairs.
When is COLM 2026, and where?
The main conference runs 6 to 8 October 2026 at the Hilton Union Square in San Francisco, with workshops on 9 October. Decisions went out on 8 July 2026. If COLM keeps its three-edition pattern of late-March full-paper deadlines, the next submission window closes in late March 2027.
When is the COLM camera-ready deadline?
COLM does not publish a camera-ready date on its public dates page; accepted authors receive it through OpenReview. What COLM does fix, across its FAQ and its call for papers, is the format: one extra page of main text at camera-ready, and an author list frozen at the submission deadline that no camera-ready revision may change.
Free tool, no account
Where should this paper go?
Paste the title and the abstract. It ranks the venues that publish this kind of work, with the reach and the article charge OpenAlex has on record.
Take the trade, or find another room
COLM's chairs put both halves of it in writing: unranked, and no easier. If that trade does not suit this paper, the matcher ranks the venues that publish work like it, free and with no account.
Matching needs no account. A paid review does not predict a decision, and what it cannot do is written down. For what kills a paper before review, see desk rejection explained. For what kills it afterwards, why papers get rejected.
phdflow reads a finished draft against the guidelines of the venue you are aiming at and returns an accept or reject call with the objections behind it. Scored against 697 real decisions: AUC 0.91 and 96% accuracy on 297 ICLR 2025 papers, one false accept in 149 rejects. The full measurement. Read a finished review before you pay for one. Your first paper gets a free preview; the full report is 60 credits, €9. No account needed. Pricing.