ICML 2027 Contribution Types

The ICML community benefits from a broad range of scientific contributions. Important advances in machine learning may take the form of new methods, theoretical understanding, insights arising from real world applications, ambitious new concepts, informative negative results, or new datasets, evaluations, and benchmarks.

ICML 2027 requires every submission to select a Contribution Type that best reflects the paper's primary scientific contribution. The selected Contribution Type will guide how the paper is reviewed and could be used as one factor in reviewer and Area Chair matching.

Required Contribution Type

All submissions must select exactly one Contribution Type at submission.

The available Contribution Types are

1. General
2. Theory
3. Application-driven
4. Concept and Feasibility
5. Negative Results
6. Datasets, Evaluations, and Benchmarks

The selected Contribution Type should describe the paper's primary scientific contribution, not every contribution contained in the paper.

Many papers will naturally contain elements of several Contribution Types. Authors should select the type that best represents the main contribution they want reviewers to evaluate.

Contribution Types are intended to guide scientific evaluation. They are not mutually exclusive descriptions of what a paper may contain.

General

We expect many ICML submissions to fall into the General Contribution Type.

General papers make contributions that are naturally evaluated using the standard expectations for machine learning research and do not fit more naturally within one of the specialized Contribution Types.

Such contributions may include new methods, algorithms, empirical findings, conceptual insights, systems, applications, or combinations of these.

Theory

Select Theory when the primary contribution consists of theoretical analysis, mathematical results, new theoretical frameworks, or proofs.

Theoretical work does not need to outperform state of the art empirical methods in order to be significant.

Empirical validation is not required when the theoretical results stand on their own.

However, this does not mean that empirical evidence is excluded from Theory contributions. Authors may include experiments to illustrate theoretical phenomena, validate predictions, clarify assumptions, or demonstrate relevance beyond the exact theoretical setting.

Reviewers may also reasonably request appropriately scoped empirical validation, including experiments on simple, synthetic, or controlled models, when such evidence would materially help validate or interpret the claims of the paper.

A Theory submission should not be required to conduct extensive empirical studies merely because such studies would normally be expected for a primarily empirical contribution.

Application-driven

Select Application-driven when the primary contribution arises from addressing the needs of a meaningful real world application or use case.

The use case should substantially inform the problem formulation, methodology, data, constraints, evaluation, or other scientific choices in the paper.

An application-driven contribution may be original because it introduces a new problem formulation, combines existing methods in a novel way, incorporates important domain knowledge, or provides new insights into an important real world problem.

Merely applying an existing machine learning method to a new dataset would generally not be sufficient unless the framing or analysis is novel.

Concept and Feasibility

Select Concept and Feasibility when the primary contribution is a highly novel idea with potentially broad significance whose full scope cannot reasonably be validated within a single paper.

The paper must still provide convincing evidence that the central concept is viable.

This Contribution Type is not intended for preliminary or incomplete work.

The originality and significance bar for Concept and Feasibility contributions is high.

Negative Results

Select Negative Results when the primary contribution is a rigorous and informative negative finding that changes our understanding of an important question.

The contribution may expose an important limitation, challenge a widely held assumption, demonstrate that a promising approach fails under meaningful conditions, or establish that a desired result is impossible or unlikely.

A failed experiment alone is not sufficient.

The paper should provide analysis, evidence, or theory that explains why the negative result matters and what can be learned from it.

The originality and significance bar for Negative Results contributions is high.

Datasets, Evaluations and Benchmarks

Select Datasets, Evaluations and Benchmarks when the primary contribution advances the resources, methods, or scientific understanding used to evaluate machine learning systems.

Examples include

1. introducing a new dataset or data resource that enables important scientific questions
2. introducing a new benchmark or substantially redesigning an existing benchmark
3. developing new evaluation methodologies or metrics
4. developing evaluation tools, frameworks, or infrastructure
5. auditing or stress testing machine learning systems
6. identifying limitations, contamination, saturation, bias, or failure modes in widely used benchmarks or evaluations
7. developing human centered or interaction based evaluation methodologies
8. conducting substantive data centric or evaluation focused empirical analyses that lead to new scientific understanding

A Datasets, Evaluations and Benchmarks contribution does not need to introduce a new learning algorithm.

Originality may instead arise from a new evaluation question, metric, protocol, benchmark design, dataset construction methodology, auditing technique, or from important new insights revealed through rigorous evaluation of existing systems.

A new dataset or benchmark should not be evaluated primarily according to whether it enables a new state of the art result. Its contribution may instead lie in enabling new questions, revealing previously inaccessible phenomena, supporting more reliable comparisons, or improving how progress is measured.

Authors should clearly state what scientific claims the dataset, benchmark, or evaluation supports and explain the assumptions and limitations under which those claims hold.

When relevant, authors should address issues such as licensing, privacy, consent, data provenance, annotation practices, potential biases, contamination, intended use, and limitations.

Contribution Types Are Not Exclusive

The Contribution Type identifies the primary basis on which the paper should be evaluated. It does not define everything that may appear in the paper.

A Theory paper may include substantial experiments.

An Application-Driven paper may contain important theoretical results.

A Datasets, Evaluations and Benchmarks paper may introduce methodological innovations.

A Concept and Feasibility paper may contain substantial empirical validation.

A Negative Results paper may propose new methods or directions motivated by the negative finding.

Reviewers should evaluate secondary contributions using appropriate scientific standards while keeping the primary Contribution Type in view.

Requests for Additional Evidence

A paper should not be penalized for lacking a form of evidence that is not necessary to establish its primary contribution.

At the same time, reviewers may reasonably request additional experiments, theory, analysis, comparisons, or other evidence when these are needed to validate, clarify, or strengthen claims made by the paper.

Reviewers should distinguish between

1. evidence necessary to support the central claims of the paper
2. evidence that would strengthen the paper but is not essential
3. evidence that would effectively require the authors to turn the submission into a different type of contribution

The first may appropriately affect the evaluation of the submission.

The second may be suggested as an improvement.

The third should generally not be required.

The Contribution Type should guide what constitutes a sufficient primary contribution without preventing reviewers from considering other forms of evidence that are relevant to the paper's claims.

Contribution Types and Reviewing

Reviewers and Area Chairs will be asked to indicate the Contribution Types they are comfortable evaluating.

Contribution Type preferences could be used as one weighted factor in reviewer and Area Chair matching together with subject matter expertise and other matching signals.

Reviewers must identify the Contribution Type selected by the authors and read the corresponding ICML 2027 reviewing guidance before completing their review.

The review form may differ across Contribution Types in order to reflect the standards most appropriate to each form of scientific contribution.

Reviewers should evaluate the submission according to the Contribution Type selected by the authors rather than substituting a different type based on their own preference.

Contribution Type Confirmation

The review form will ask reviewers to confirm the Contribution Type selected by the authors and that they have read the corresponding reviewing guidance.

The reviewer selected Contribution Type must match the type listed in the submission.

ICML may automatically check for mismatches between the Contribution Type selected by the authors and the Contribution Type entered by reviewers.

If there is a mismatch, reviewers may be asked to confirm the correct Contribution Type, read the corresponding guidance, and revisit their review as necessary.

The purpose of this check is to ensure that each submission is evaluated according to the appropriate reviewing standards.

Contribution Types and Acceptance

Contribution Types are not separate tracks and are not intended to establish separate acceptance quotas.

All Contribution Types are part of the ICML scientific program and should be held to a high standard of scientific quality.

Different Contribution Types may demonstrate scientific excellence in different ways. The purpose of the Contribution Type system is to ensure that these differences are recognized consistently and fairly.