Die strengsten Regeln in der Geschichte der ICML 2026: LLMs dürfen nicht als Autoren aufgeführt werden, und Manuskripte, bei denen es zu einem Missbrauch von KI kommt, werden direkt abgelehnt.
The call for papers for ICML 2026 is officially open!
Today, the ICML organizing committee released relevant notes for submissions to this top - tier conference. Three points are of crucial importance:
Authors of accepted papers can freely choose whether to attend the conference.
Release the "original submitted version" of accepted papers.
Set an upper limit on the number of reciprocal review papers.
In addition, ICML 2026 has made it clear that the use of LLMs to assist in writing or research is allowed, but all responsibilities rest with the authors.
This includes content generated by AI that may be considered plagiarism or academic misconduct.
Meanwhile, the organizing committee prohibits listing LLMs as authors and strictly forbids "prompt injection"; otherwise, the submission will be directly rejected.
As the 43rd International Conference on Machine Learning, ICML 2026 will be held in Seoul, South Korea, from July 7th to 12th, 2026.
Deadline for full paper submission: January 28th, 2026.
ICML Paper Call, Fully Open
According to the official announcement, all papers will be reviewed through a double - blind review process, and accepted papers will be presented at the conference.
Papers must be prepared and submitted as a single file: 8 pages for the main paper, with no page limit for references, impact statements, and appendices.
There is no separate deadline for submitting supplementary materials. The final version of accepted papers will allow an additional page for the main paper.
Next, here are the policy changes of ICML 2026 that are different from the past:
1. Attendance of accepted paper authors is optional
After receiving the acceptance notice, authors can decide before the specified date whether they wish to present their papers in person at the conference or only have their papers included in the proceedings (without presenting at the conference).
Regardless of the choice, all accepted papers will be treated equally in the proceedings.
They will all be eligible for ICML awards and the outstanding titles equivalent to the past "oral presentations" and "spotlight posters".
For papers only included in the proceedings, at least one author must register virtually.
2. Release the original submitted version and the camera - ready version for accepted papers
For all accepted papers, in addition to the camera - ready version, the following materials will also be released: the original submitted version (including supplementary materials), anonymous review comments, meta - review comments, rebuttals, and discussions between reviewers and authors.
Authors of rejected papers will also have the option to release their original submitted versions, review comments, meta - review comments, rebuttals, and discussions between reviewers and authors.
3. Set an upper limit on the number of reciprocal review papers
In addition, ICML sets an upper limit on the number of papers for which a person can be designated as a reciprocal reviewer and updates the policy on the use of AI tools in the review process.
Of course, ICML 2026 allows a wide range of topics for paper submissions, including but not limited to:
General machine learning (active learning, clustering, online learning, ranking, supervised, semi - supervised and self - supervised learning, time series analysis, etc.)
Deep learning (architectures, generative models, theory, etc.)
Evaluation (methodology, meta - research, reproducibility and validity, human - machine interaction, etc.)
Machine learning theory (statistical learning theory, bandits, game theory, decision theory, etc.)
Machine learning systems (improving implementation and scalability, hardware, libraries, distributed methods, etc.)
Optimization (convex and non - convex optimization, matrix/tensor methods, stochastic, online, non - smooth, composite, etc.)
Probabilistic methods (Bayesian methods, graphical models, Monte Carlo methods, etc.)
Reinforcement learning (decision - making and control, planning, hierarchical reinforcement learning, robotics, etc.)
Trustworthy machine learning (reliability, causality, fairness, interpretability, privacy, robustness, security, etc.)
Application - driven machine learning
Next, here are the key points, important dates, and submission sites:
Submission site opens: January 8th, 2026.
Abstract submission deadline: January 23rd, 2026, AoE (noon, January 24th, 2026, UTC - 0).
Full paper submission deadline: January 28th, 2026, AoE (noon, January 29th, 2026, UTC - 0).
Abstracts and papers can be submitted via OpenReview: https://openreview.net/group?id=ICML.cc/2026/Conference
Note: Position papers should be submitted via a separate OpenReview site, as described in the call for position papers.
AI can assist in writing, but cannot be listed as an author
In the policy, the regulations on the use of generative AI are particularly crucial.
This time, ICML 2026 allows authors to use GenAI tools, such as large models, to assist in writing or research.
However, authors must take full responsibility for all content in their papers, including any content generated by AI tools that may be considered plagiarism or scientific misconduct.
LLMs do not qualify as authors;
Any "prompt injection" is strictly prohibited, and the submission will be directly rejected. Prompt injection refers to inserting invisible text in the paper with the intention of manipulating the LLM to obtain a favorable review;
ICML is considering various options for incorporating AI tools (such as LLMs) in the review process. In addition, it is not allowed to fully delegate the review to AI.
Regarding reciprocal review, eligible authors must review papers for ICML. However, there are two requirements: per - submission requirement and per - author requirement.
Per - submission requirement:
All submissions must have at least one author agree to serve as an ICML reviewer.
The designated author should meet the definition of reviewer eligibility given in the peer - review FAQ. A new rule this year: An author can be designated as a reciprocal reviewer for a maximum of 2 submissions of which they are an author.
Exception: If no author is eligible (as defined in the peer - review FAQ), or all eligible authors are already serving as reciprocal reviewers for their 2 submissions, or are serving in SAC, AC, or other organizational roles for ICML 2026, the submission is exempt from this requirement.
To meet this requirement or claim an exception: The abstract submission form will allow submitters to designate an author to meet this requirement or indicate that the submission is exempt from this requirement.
Per - author requirement:
In addition, each author with 4 or more submissions must agree to serve as an ICML reviewer (in this case, the author can serve as a reciprocal reviewer for a maximum of 2 of their submissions).
If there is an unforeseen shortage of eligible reviewers, the program chair may lower the threshold to 3 submissions (if this happens, authors newly subject to this requirement will be notified).
Exception: If an author is serving in AC, SAC, or other organizational roles for ICML 2026, they are exempt from this requirement.
To meet this requirement or claim an exception: Authors with 4 or more submissions should fill out the per - author reciprocal review form (link to be announced) to provide the information we require from all reviewers or indicate that they are exempt from this requirement.
As in the past, all submissions must be anonymized in the double - blind review process and must not contain any information that intends or has the consequence of violating the double - blind review policy.
Authors are allowed to post versions of their work on preprint servers such as arXiv. They are also allowed to give talks about their submitted work during the review period.
However, during the review period, i.e., from the time the author submits the paper to the time the acceptance/rejection decision is communicated, the work should not be promoted as an ICML submission under any circumstances.
If an author has published or plans to publish a non - anonymous version of the paper before the ICML decision is made, the submitted version must not reference the non - anonymous version.
Authors must strictly follow the standard ethical behavior of peer review. In particular:
Any form of plagiarism is prohibited.
Prompt injection is prohibited.
Promoting the work as submitted to ICML during the review period (such as in talks or on social media) is prohibited.
Any form of collusion, whether explicit or tacit (authors collaborating with reviewers, ACs, or SACs to obtain favorable or unfavorable reviews) is prohibited.
In addition, authors must not submit papers that are the same or substantially similar to versions that have been published, accepted for publication, or concurrently submitted to other conferences or journals.
Such submissions violate our double - submission policy, and the organizers have the right to reject such submissions or remove them from the proceedings.
Any concurrent ICML submissions with overlapping author sets will also be considered prior work. For example, if publishing one would make the other too incremental, this may be considered a reason for rejection.
The review criteria require that submissions should report original and rigorous research of significant importance to the machine - learning community.
All claims must be clearly stated and supported by reproducible experiments and/or reliable theoretical analysis.
Contributions must be placed in the context of the broader scientific and machine - learning research literature, with appropriate acknowledgment and distinction of relevant prior work.
References:
https://x.com/icmlconf/status/1986089104367308805
https://icml.cc/Conferences/2026/CallForPapers
This article is from the WeChat official account "New Intelligence Yuan". Author: New Intelligence Yuan. Editor: Taozi. Republished by 36Kr with permission.