Colt camera ready

COLT Camera Ready

Use this after a COLT acceptance. COLT proceedings are published electronically in the Proceedings of Machine Learning Research (PMLR) — verified against the COLT 2026 CFP and PMLR volume pages on 2026-07-08. The camera-ready instructions arrive with the acceptance notification; they, not this file, control exact dates and forms.

Recent volume anchors (verified 2026-07-08)

Edition	PMLR volume	Location	Volume editors
COLT 2024 (37th)	v247	Edmonton, Canada	Shipra Agrawal, Aaron Roth
COLT 2025 (38th)	v291	Lyon, France	Nika Haghtalab, Ankur Moitra
COLT 2026 (39th)	announced with acceptance (待核实)	San Diego, USA	current program chairs

Volume editors are the program chairs of that edition; the pairing rotates every year. The COLT 2026 volume number and camera-ready deadline were not yet published at the access date — take both from your acceptance email.

De-anonymization pass

• Remove the template's anon option so the author block renders; verify the running head and PDF metadata now carry real names.
• Restore acknowledgements, funding lines, and any statements the anonymous version cut. These lines add length — recheck where the body boundary falls if the current instructions still enforce a body limit at camera-ready (待核实 per cycle).
• Rewrite defensive third-person self-citations ("the authors of [12] showed") back into honest first person where it reads better.
• Swap anonymized or omitted repository links for stable public URLs, and test each from a logged-out browser.

Consolidating the archival PDF

At COLT the appendix ships inside the same archival PDF, so the proceedings version is the complete record — there is no separately hosted supplement to lean on later.

• Fold every rebuttal-promised edit into the text now: theorem statements that must carry an explicit constant, assumptions promoted from appendix to body, corrected displays. Keep a private ledger mapping each reviewer request to the diff that answers it, because chairs occasionally spot-check.
• Do not renumber theorems, lemmas, or assumptions relative to the reviewed version unless forced; the review record and any citing preprints refer to the old numbers.
• Reconcile the arXiv version: theory readers will diff arXiv against PMLR, so either update arXiv to match or note the differences in the arXiv comments field.
• Run a final pass on cross-references — \ref breakage introduced by de-anonymization edits is the most common camera-ready defect in proof-heavy papers.

PMLR metadata and citation hygiene

Title, author order, and abstract you enter in the proceedings forms become the citable record. The PMLR page for your paper generates the canonical citation; make your own BibTeX consistent with it:

@inproceedings{yourkey2026tight,
  title     = {Tight Rates for a Learning Problem},
  author    = {Lastname, First and Coauthor, Second},
  booktitle = {Proceedings of the 39th Annual Conference on Learning Theory},
  series    = {Proceedings of Machine Learning Research},
  volume    = {TBD},   % assigned volume number, e.g. v291 was COLT 2025
  pages     = {1--30}, % assigned by the volume
  year      = {2026},
  publisher = {PMLR}
}

• Check every accent and math symbol in the title as it will render on the PMLR page.
• Author order must match the submission unless the chairs approved a change.
• PMLR is open access with no author-side publication fee; COLT's cost model is conference registration, not an APC.

Presentation and registration obligations

• Expect that an accepted COLT paper must be presented at the conference by an author; the exact registration requirement, talk versus poster format, and any remote option for the 2026 edition were not verifiable at the access date (待核实) — the acceptance email and the registration page at learningtheory.org control.
• COLT is single-track: your talk or poster is seen by the whole community, so slide preparation deserves more lead time than at parallel-session venues.
• Budget for registration early; COLT 2026 ran June 29 to July 3 in San Diego, and conference-hotel capacity at smaller venues fills quickly.

Preparing the talk alongside the PDF

Camera-ready season doubles as talk season, and at a single-track venue the talk is part of the paper's public record:

• Build the talk from the informal theorem statements and the technique paragraph, not from the formal setup — the room already speaks the field's vocabulary but has not seen your notation.
• One slide of "what was known" with exact prior rates mirrors the paper's known-vs-new ledger and is the single highest-value slide at COLT.
• Rehearse the proof-idea slide until it survives without the appendix: the aim is that a listener could set a student to re-derive the result.
• If the edition uses posters for some papers, the poster's spine is the same three objects: model, theorem, technique — resist the empirical-venue urge to fill space with plots you do not have.
• Announce the open questions from the paper's discussion section; COLT audiences treat posed problems as gifts, and follow-up work cites the talk's source paper.

Common camera-ready defects at COLT

Defect	Why it happens	Catch it by
Body overflow after author block returns	Anonymous version used the freed lines	Recompiling with authors before editing anything
Theorem numbering drift vs. reviewed version	Appendix restructuring	Diffing the numbered-statement list
arXiv/PMLR content divergence	Only one copy updated	A scheduled same-day update of both
Broken \ref after merging rebuttal edits	Labels deleted with old text	grep-style check for "??" in the compiled log
Metadata mismatch (title case, accents)	Forms filled from memory	Copy-pasting from the final PDF

Output format

[Camera-ready status] Ready / Needs fixes / Blocked
[Reviewer-promise ledger] <request -> diff location>
[De-anonymization checks] <author block / acknowledgements / links / metadata>
[Numbering stability] preserved / documented changes
[Logistics] <forms, registration, presentation owner>