colm-topic-selection

COLM Topic Selection

Almost every LM paper could be sent to five venues; COLM exists because the authors

of such papers kept finding that none of the five was actually about their question.

This skill decides whether your project is one of those papers. Run it before the

first draft — the March abstract deadline punishes late routing changes.

The object-of-study test

Ask one question: **is the language model the thing being studied, or the tool doing

the studying?**

• Studied — you train, measure, dissect, steer, attack, or critique LMs, and the

finding is about LMs themselves. That is COLM's declared identity: a venue for

understanding, improving, and critiquing LM technology, created (announced October

2023, first edition 2024) because these questions sat awkwardly at every older

venue.

• Tool — the LM is fixed infrastructure and the contribution lives in a task,

domain, or application. Route to the task's home venue; COLM reviewers will ask

what was learned about language models and find nothing.

A useful second probe: delete the specific model names from your abstract. If the

claims survive as statements about language modeling (a training dynamic, a data

effect, an evaluation artifact, an inference trade-off), COLM fits. If what remains

is a task result, it does not.

Routing against the adjacent venues

| Your project's center of gravity | Better home | Why not COLM |

|---|---|---|

| A training method, data effect, evaluation artifact, inference algorithm, or safety property of LMs | COLM | — |

| A linguistic phenomenon, an NLP task, an annotation resource, a multilingual application | ACL or EMNLP | The finding is about language or a task, not about the model class |

| A representation-learning idea tested beyond language (vision, RL, graphs) | ICLR | Generality is the point; the LM is one instance |

| A learning-theoretic result, an optimizer, a probabilistic method where LMs are one benchmark | NeurIPS or ICML | The LM is an experiment, not the subject |

| A serving/systems contribution (kernels, scheduling, memory) with modeling untouched | MLSys or a systems venue | Reviewers there can evaluate throughput claims properly |

| An early-stage probe, negative result, or position piece not ready for archival review | A COLM or *CL workshop | Archival COLM review will demand completeness the work does not have yet |

Boundary cases worth naming: multimodal models fit COLM when the language model

is central to the question (a 2025 Outstanding Paper is a VLM-failure analysis);

agents and tool use fit when the paper studies the LM's behavior in the loop

rather than ships a product; benchmark papers fit when the benchmark comes with a

validity argument, not just difficulty.

COLM's declared lanes (2026 CFP, checked 2026-07-08)

The CFP's non-exhaustive topics cluster into lanes; name yours before framing:

• Training — fine-tuning, instruction tuning, reinforcement learning, prompt

tuning, in-context learning.

• Data — corpora and curation for pre-training, post-training, every stage.
• Evaluation — static and dynamic benchmarks, simulation environments, scalable

oversight, protocols and metrics, human and machine evaluation, bias/equity/misuse

measurement.

• Safety — security, privacy, misinformation, adversarial attacks and defenses.
• Plus inference/generation, interpretability, and multimodal LM work, all visible in

the accepted lists of 2024-2025.

The award lineage (see resources/exemplars/library.md) tilts toward measurement,

analysis, and evaluation critique: five of the eight Outstanding Papers from the

first two editions study how we know what LMs do. A single leaderboard table is a

weak spine here.

The young-venue calculus

Choosing a three-edition-old conference is itself a decision. Weigh it explicitly:

FOR COLM                                  AGAINST (for this project)
- reviewer pool self-selected for LM      - no long citation tail yet; some
  work; less "why not test on vision?"      committees still ask "is COLM top-tier?"
- award lineage rewards analysis and      - norms shift per edition (policies,
  measurement, not only SOTA                dates, format have all moved 2024→2026)
- one cycle per year, October venue,      - one cycle per year: a reject costs 12
  proceedings free and open on              months at this venue
  OpenReview

The single-cycle point cuts both ways: COLM's March-to-July pipeline sits neatly

between the ICLR (autumn) and NeurIPS (spring) deadlines, which is why many groups

now treat it as the natural resubmission target for strong-but-rejected LM papers —

plan the fallback chain, not just the first shot.

Vignette: one project, three venues

A group builds a retrieval-augmented clinical-notes summarizer and observes that

retrieved passages change the model's factuality unevenly across note types. Three

papers hide in this project:

• The system — architecture, deployment, clinician study → a clinical-NLP or

applications track (ACL/EMNLP territory).

• The task result — new SOTA on a summarization benchmark → weak everywhere as a

sole contribution, and weakest at COLM.

• The phenomenon — retrieval shifts factuality via an identifiable mechanism,

measured across model families with contamination-controlled evaluation → this

is the COLM paper, and notice the clinical domain has become incidental.

The exercise generalizes: the COLM-shaped paper inside a project is usually found

by asking what you now know about language models that you did not know before the

project — then testing whether that knowledge survives on models you did not build

the system around.

Commit checklist before writing

• State the one-sentence finding about language models that survives model-name

deletion.

• Name the CFP lane and the two nearest COLM accepted papers (2024 or 2025 lists)

your work will be shelved next to.

• Confirm the evidence plan can clear the venue's measurement bar: pinned models,

contamination discussion, uncertainty over runs (colm-experiments).

• Confirm the project can be told in 9 pages of main text — the 2026 cap is strict.
• Check the current CFP; COLM's scope wording is young enough to move between

editions.

Output format

[Routing] COLM / ACL-EMNLP / ICLR / NeurIPS-ICML / systems venue / workshop first
[Object-of-study test] passes / fails — <what the LM is in this project>
[CFP lane] training / data / evaluation / inference / safety / interpretability
[Nearest COLM neighbors] <two accepted papers>
[Young-venue risk accepted] yes / no — <one-line reason>
[Next action] <framing fix, evidence fix, or venue switch>