COLM Submission

Everything below is the COLM 2026 cycle as read from colmweb.org on 2026-07-08. COLM re-issues its instructions each edition and has already changed details between 2024, 2025, and 2026 — reopen the current submission-instructions page before acting on any number here.

The 2026 mechanics

• Two deadlines a week apart: abstract March 26, 2026; full paper March 31, 2026. The abstract is not decorative — it routes reviewer bidding — so submit the real abstract, not a stub you plan to overwrite.
• Platform: OpenReview, group colmweb.org/COLM/2026/Conference, opened February 2026. Every author needs an OpenReview profile with current affiliation and publication history before the deadline; profile gaps break conflict detection and can strand the submission.
• Format: a strict 9-page upper limit on main text, with unlimited additional pages for citations. An optional acknowledgments section (≤ 1 page) sits outside the limit — but in the submitted version there must be no acknowledgments at all, because review is double-blind.
• Anonymity: no author names, no acknowledgments, and no links that resolve to you — the instructions call out GitHub links explicitly. LM papers leak identity through unusual channels; sweep them (next section).
• Code of Ethics: every author must read it and explicitly acknowledge it in the submission form. Reviewers are encouraged to flag violations, so treat the acknowledgment as a real audit, not a checkbox.
• LLM-usage disclosure: COLM 2026 adopted the ICLR 2026 LLM policy minus minor writing/code assistance. Disclose LLM-originated ideas, LLM-written content (including references), LLM-generated data or plots, and LLM-based evaluation. Fill this in truthfully; a judge-model evaluation pipeline is disclosable use.
• Reciprocal reviewing: there is a per-submission and a per-reviewer requirement; an author with 4+ submissions enters the reviewer pool automatically, and two submissions cannot lean on the same reciprocal reviewer unless nobody else qualifies. Decide the nominee per paper before deadline night — group-wide, this is a constraint-satisfaction problem, not a form field.

Where LM papers leak identity

Beyond the obvious author block: Hugging Face model and dataset URLs under your org, Weights & Biases dashboard links, API keys or org IDs left in config listings, model names that only your lab uses internally, "our previously released model X" phrasing, dataset cards naming maintainers, and PDF metadata from your LaTeX toolchain. Grep before upload:

# Anonymity + hygiene sweep over the built PDF and the source tree
pdftotext main.pdf - | grep -niE 'huggingface\.co/[a-z0-9-]+|wandb\.ai|github\.com|we (previously|earlier) released|acknowledg' | head
pdfinfo main.pdf | grep -iE 'author|creator|producer'
grep -rniE 'api[_-]?key|org-[A-Za-z0-9]{6,}|hf_[A-Za-z0-9]{20,}' paper-src/ | head

Every hit is either a required anonymization fix or a leaked credential — both are blocking.

Risk triage before upload

Deadline-week order of operations

1. T-7 days: freeze the claim set; anything unverified by now gets cut or hedged, not rushed.
2. T-5: submit the real abstract (March 26 in 2026). Confirm all OpenReview profiles and the reciprocal-reviewer assignment.
3. T-3: assemble the full PDF on the official COLM template; run the sweep commands above on the built PDF.
4. T-2: collect every co-author's Code of Ethics acknowledgment and LLM-usage disclosure input; these require other humans, so they cannot be T-0 tasks.
5. T-0 (March 31): upload, then re-download your own submission and read it cold — what you download is what reviewers receive.

The abstract week is a strategy window

The five days between abstract (March 26) and paper (March 31) in 2026 were not slack — they are the last point where cheap decisions are still available:

• Title and abstract drive reviewer bidding. Include the load-bearing nouns a matching system and a bidding human need: the phenomenon, the method family, the model classes studied. A cute title with no searchable terms buys you the wrong reviewers for four months.
• Subject-area selections route the paper. Choose the lane where your weakest section will be judged most fairly — an evaluation-critique paper mis-filed under "training methods" gets reviewers who ask for training contributions.
• The withdraw option is real. If the claims ledger still has red entries at the abstract deadline, submitting the abstract costs nothing, and withdrawing before March 31 costs only pride. Submitting a paper whose headline experiment finished an hour before upload costs a year at this venue.

Abstract-week triage:
  all claims verified            -> proceed; polish
  1-2 claims at risk             -> submit abstract; cut/hedge at-risk claims now
  core claim unverified          -> submit abstract; decide by T-2, not T-0
  core claim failed              -> withdraw; COLM 2027 or the autumn venue

Cycle-volatile items — recheck each edition

• All dates (the March 26 / March 31 pair is a 2026 fact, not a pattern; 2024's deadline was in late March too, but COLM has not promised stability).
• The 9-page cap, acknowledgment allowance, and template version.
• LLM-policy wording — 2026's is one edition old and explicitly adapted from another venue's; expect revision.
• Reciprocal-reviewing thresholds and mechanics.
• 待核实: any supplementary-material size caps and any abstract-edit grace rules; neither was verifiable on 2026-07-08.

Output format

[COLM submission status] ready / blocked / at risk
[Format] page count vs 9-page cap; template OK?
[Anonymity] clean / leaks: <channel list>
[Policy items] ethics-ack ▢  llm-disclosure ▢  reciprocal-reviewer ▢
[Fix queue] <ordered, with owners — co-author items first>