ICLR 2027

ICLR Program Chairs: Jacob Andreas, Guy Van den Broeck, Andrej Risteski, Stella Yu

Google Research: Rajesh Jayaram, Vincent Cohen-Addad, Drew Tyler, David Woodruff, Tommaso D’Orsi

Google Cloud: Jinsung Yoon, Mihir Parmar, Palash Goyal

Google Sponsors: Corinna Cortes, Vahab Mirrokni, Tomas Pfister, Burak Gokturk

Following positive feedback from other venues, like STOC, ICML and NeurIPS, we are pleased to announce a new initiative for ICLR 2027, in partnership with Google, that will provide authors access to their Paper Assistant Tool (PAT) and support authors in improving their submissions. 

This program offers authors a limited opportunity to receive free, automated, and actionable feedback on their manuscripts from September 11 to September 18, privately to the authors. The feedback the authors will receive from PAT through the ICLR program will not be used in the review process. Reviewers, area chairs, and program committee members will not have access to the PAT feedback. 

Our goal is to broaden access to capable AI models so that all ICLR authors can benefit equally from these tools—and, in doing so, help raise the overall standard of submissions.

What is PAT?

PAT is a specialized, experimental tool powered by Google’s Gemini models, utilizing a “reasoning”-focused pipeline to offer AI-powered feedback on computer science papers. It is similar to those that have achieved high-level performance on mathematical problem-solving benchmarks. The model is designed to help authors identify issues that human reviewers might flag, including (but not limited to) experimental and methodological rigor, narrative clarity in English, and technical correctness.

In a pilot at the Annual ACM Symposium on Theory of Computing, STOC, (blog), 94% of participants found the pre-submission feedback generated by an AI assistant to be helpful, and 85% reported that the feedback resulted in improved clarity of their paper. Following this pilot, PAT was expanded and launched in partnership with ICML (blog), which had a similarly positive reception. Notably, 35.4% of responding authors with theoretical results reported the tool identified significant theory gaps that took more than an hour to fix, and 31% of responding authors with experimental results said the feedback prompted them to run new experiments. An updated version of PAT was subsequently deployed at NeurIPS (blog), where 54.4% of respondents reported an improvement in performance.

While the STOC program focused heavily on theoretical correctness, the ICML and NeurIPS programs were specifically tuned to address the needs of the machine learning community, incorporating author feedback and integrating components of the ScholarPeer system into the PAT feedback. The version of PAT deployed today further leverages stronger underlying models with more advanced reasoning capabilities to provide more accurate and insightful feedback to computer scientists.

Like theNeurIPS,  ICML and STOC programs, the goal is to help authors at ICLR to improve the quality of their papers, not to replace human peer review. By fixing clarity issues and potential technical gaps before the work is submitted to the conference, we hope to give authors actionable feedback before their paper enters the review process.

Logistics and Eligibility

The program is entirely optional. It operates inside OpenReview, but completely outside the official review process. The program will run for a 7 day window, between September 11 and September 18 (11:59 pm, Anywhere on Earth). 

To manage resources fairly, each eligible author is granted one virtual “voucher” to have a single paper run through the AI feedback system. In addition, each paper can be run through the system at most once. Any author who has a valid OpenReview Account is eligible to use the system.

To redeem this voucher, authors will select a checkbox “Ready for LLM Feedback” on the OpenReview paper submission form to flag the manuscript for AI review. This feature will only work after a PDF has been successfully uploaded to the OpenReview server. The author who checks the box redeems their voucher for the paper once the edit to the submission is submitted, and the version of the document at the time of the edit is then sent to the pipeline for automated feedback. If an ineligible author or an author who has already used their voucher attempts to select the “Ready for LLM Feedback” button, an error message will appear and the paper will not be sent out for review. Eligible submissions will typically receive the feedback within 12 hours of being submitted. 

Submission Timing: To ensure system stability and incentivize early submissions, papers submitted earlier in the feedback window are guaranteed the full compute budget of the PAT pipeline. Submissions made very close to the deadline (1-2 days) may be subject to throttling of their overall compute allocation depending on the demand. 

The technical staff will be able to provide limited, best-effort support on the program, such as answering questions or checking for failed paper delivery. Please direct such questions, as well as any feedback you may have on the program, to paper-assistant@google.com.  

Privacy and Data Safety

We recognize the sensitivity of unpublished research. Trust is the cornerstone of this experiment, and we have implemented strict protocols to ensure author safety:

1. Strict Separation from Peer Review: The AI Feedback is entirely independent of the ICLR review process. It is visible only to the authors in OpenReview. Reviewers, Area Chairs, and Program Chairs will have no access to this feedback. Furthermore, the PAT system will not be used in any part of the ICLR review process. 
2. Stateless Inference (No Training): Submissions will not be used to train, fine-tune, or improve Google’s models. The model operates in a stateless “inference-only” mode; it processes the text to generate feedback and retains no memory of the specific content for future learning.
3. Data Destruction: To minimize data exposure, Google will employ a strict deletion policy. All PDFs and feedback submitted to Google are stored in a restricted access environment and are scheduled for permanent deletion within 7 days after the feedback is delivered and the program is completed.
4. Restricted Access: Google staff who require access for product implementation will have access to the data (submission PDFs and generated feedback). They will only do so in the event of a technical difficulty, and only after explicit author approval.

Caveats and Disclaimers

Like all LLMs, the models used by the PAT pipeline are not infallible. Authors should treat the generated feedback with the same critical eye they would apply to any other LLM or a human review. In particular:

• The model may occasionally flag correct statements as errors, as well as miss actual flaws. It is the author’s responsibility to verify the validity of the feedback.
• The model may make suggestions for the paper that you disagree with. By considering why you disagree with the suggestion, you may be able to add clarifications that would help other readers of the paper, including reviewers. 

Outcomes

Our primary objective is to broaden access to capable AI models so that all authors at ICLR can benefit equally from these tools. At the same time, this presents a significant opportunity to raise the standard of our own scientific submissions. In particular, PAT can help authors catch errors—in proofs, experimental setups, or structural reasoning—that might otherwise be missed by both authors and reviewers.

After the full paper deadline, an anonymous author survey will be sent to authors who used PAT to request feedback so that Google can improve PAT. Submitting feedback via the survey is optional. We look forward to seeing the results of the program, which will be shared with the broader community via the ICLR blog once the program is over.

FAQ

PDF Size: Due to context limitations, very large PDF’s containing, for example, multiple high resolution images or plots, may have their images stripped before processing. This would result in PAT running only on the extracted text portion of the paper. In extreme cases, the pipeline may fail altogether. We recommend submitting PDFs no larger than 20MB (ideally less than 10) to the PAT system to ensure success of the pipeline. 

Turnaround Time: We expect that feedback will be posted within 1-2 hours of submission to the PAT system. During times of high demand, such as closer to the deadline, the latency may be longer. We strive to have all feedback posted within 12 hours of the submission to PAT.

How do I know my paper was successfully submitted?: After clicking the “Ready for PAT Review” button, your PDF should be picked up by the PAT system within a few minutes. At that post, a private comment (visible only to authors) will be posted to your paper notifying you that the paper has entered the PAT processing queue.

Contact (for feedback and best-effort assistance): paper-assistant@google.com

ICLR Program Chairs: Jacob Andreas, Guy Van den Broeck, Andrej Risteski, Stella Yu

One of the big changes we’ve made at ICLR this year is a set of policies that effectively “rate-limit” submissions—a 20-paper limit for all authors, and a 1-paper limit for submissions by “new authors” (defined as people who have never published at a major AI / ML / CV / robotics / NLP conference). We know that community members have strong feelings (both positive and negative!) about this new policy, so we wanted to share a little bit about the thinking that went into it.

What are we trying to accomplish?

To start with the most basic question: why does ICLR put papers through peer review at all? Why subject papers to expert judgment in the first place, rather than letting anyone who wants to present set up a poster?

• The most important function of reviewing is to help the research community separate signal from noise. The volume of machine learning research taking place is far too much for any individual to review—we need some procedure for identifying which findings should rise to the level of broad attention.
  
• An additional important function is to improve the quality of work. We want to create positive incentives for people to make their findings as robust & comprehensive as possible before they’re presented to others to read. If the resulting work still has shortcomings, the review process provides a mechanism for communicating and addressing them.
  
• As AI is increasingly a subject of public and media attention, peer review also serves to communicate credibility to non-experts, who may not be able to assess the technical merit of the research itself but must still make policy, business, and even personal decisions affected by research findings.
  
• Finally, peer review has an important role in helping institutions decide how to allocate resources (jobs, tenure, grant money, compute, etc.). The true impact of a given piece of research is often visible only with years of hindsight, but the stamp of publication still provides an intermediate quality signal that is useful for many of these institutional processes.

Notably, when we review a paper we’re really judging it on two orthogonal axes: correctness (are the paper’s main claims accompanied by convincing evidence?) and significance (is the paper’s finding new, unexpected, and important?). The first criterion is mostly objective (but in practice can require a huge amount of work to verify); the second is entirely subjective, and research communities are defined by their collective taste in problems and solutions. Current peer review mechanisms have evolved over centuries to better provide both, and ICLR itself has a long track record of innovative reviewing practices—including public discussion, the ability to update manuscripts during the review phase, and the fact that even rejected papers become public (and de-anonymized) at the end of the review process.

Reviewing in the AI era

Why make changes now? The fundamental challenge with expert review is that it requires a lot of experts. Researchers at the very beginning of their careers (who we want to train and welcome into the field) may lack the expertise to identify common mistakes in experimental design, or evaluate the significance and novelty of a paper’s findings. A stable field, in which the overall number of researchers is roughly constant, should expect to reach some equilibrium state where there are enough qualified reviewers for every paper. But in a very rapidly growing field (with many more newcomers than experts) there may simply not be enough eyes to look at all the work that people want to publish.

Sure enough, machine learning research is growing:

[Figure from Kim, Lee and Lee, ICML 2025]

Much of this growth predates the existence of AI tools capable enough to play a substantive role in the research process. But now the barrier to entry for generating something that looks like a research publication has become much lower—modern AI systems can generate paper-shaped objects even for users who are unfamiliar with the related work and writing conventions for a given area, and who are unable to understand or verify any of a paper’s main claims. Even when they’re correct, current AI systems seem to have poor taste in research questions (see the COLM chairs’ post on “theoryslop” and “slopterpretability”). As academic and industry jobs become more competitive, not only newcomers but many established researchers feel pressure to use these tools to generate a much higher volume of publications than we can possibly review. (A closely related challenge is reviewers delegating their responsibilities to AI—we’ll have another blog post on this soon!)

Our approach

We seek to trade off the following objectives: (1) providing opportunities for “newcomers” (whether new in their research careers overall, or new to the field of ML), (2) incentivize high-quality work, and (3) focusing precious reviewer attention on work that is likely to make an impact.

This year our strategy for doing so involves three pieces:

The deanonymization policy

Since the beginning of ICLR’s “open review” experiment, it has always been the case that all papers become de-anonymized at the end of the review process.

While this is not a new policy, we think it’s especially valuable in the AI era—if authors are going to ask someone to put time into reviewing their paper, they should be willing to publicly associate themselves with that paper regardless of the outcome of the review process. The current mechanism allows authors to do so while still offering double-blind reviewing.

The one-limit paper for new authors

ICLR 2027 authors are permitted at most one publication in which no author is a qualified reciprocal reviewer.

Venues across the ML world have observed that much of the rapid rise in paper submission, especially in the last year, comes from papers on which no author is qualified to review. Most conferences haven’t published data about this, but at ICLR 2026 about 20% of submissions had no reciprocal reviewer (computed by comparing the author lists to the final PC list). Within this set of “no-reviewer” submissions, about 15% are sole-author papers, and about 40% have authors who submitted multiple “no-reviewer” papers. Based on communication with other venues we expect these numbers to be much higher for ICLR 2027.

Last year, a quarter of “no-reviewer” submissions were desk-rejected. Even among those papers that went to full review, they were accepted at about half the rate of submissions with authors on the program committee. Thus papers in this group are much less likely to be accepted than the average ICLR submission (and many of them clearly do involve low-quality / AI-generated content). A the same time, many were excellent papers and accepted to the conference!

In applying a rate limit to submissions from this group, our goal is not to gate-keep based on academic qualifications—we want to make sure that all papers have a chance at fair reviewing, and expect that this policy will help all authors (including first-time submitters).

The 20-paper limit for all authors

No ICLR 2026 author may appear on more than 20 submissions.

The main purpose of this mechanism is to avoid (1) excessive spamming of the conference with slop papers, and (2) people using co-authorship to circumvent the new submitter policy at a very large scale (more on this below).

We especially want to clarify that “20 papers” does not represent a reasonable number of submissions! We personally find it difficult to imagine how one author could contribute meaningfully to 20 papers, and think this number could have defensibly been chosen to be much smaller. But we are sensitive to the fact that many submitters belong to large research groups, and don’t always have control over which papers their PI expects to be a co-author on; we felt that it would potentially put junior submitters at risk to impose more aggressive constraints on short notice.

We note also that fewer than 0.2% of authors submitted more than 20 papers last year. Many of these authors had acceptance rates substantially below the conference as a whole, and no author had more than 20 papers accepted. So we hope that this policy incentivizes even prolific labs to send their most complete work to ICLR.

Tradeoffs

It is important to acknowledge that these mechanisms aren’t Pareto improvements over the old reviewing system, and we know there will be individual submitters who may be more constrained than last year (e.g. established interdisciplinary researchers that have published outside the reciprocal reviewing list, junior researchers with a backlog of high-quality papers that are being resubmitted due to low-quality reviews in previous cycles).

We also acknowledge that it creates an incentive for qualified reciprocal reviewers to peddle “gift authorships” on work they didn’t contribute to. We are actively monitoring for evidence of such arrangements, and if you learn of one we encourage you to send an email to program-chairs@iclr.cc. We trust that if you are a senior researcher who is approached with such a request, you will not agree to it.

Our current belief is that all these costs are outweighed by the burden that will be imposed on the vast majority of submitters under the status quo, and that these measures the quality of the program and the review experience for both new and established authors.

What’s next

We view these changes as part of a long history of experimental reviewing practices at ICLR. We are also glad to see other venues trying different experimental policies—e.g. AAAI’s 2-stage process and the numerous new mechanisms being pursued by TMLR. We plan to share data about the outcomes of these experiments with other conferences (and since ICLR review data is all public, encourage you to perform your own analyses and do the same!). We expect that policies will continue to change for future iterations of the conference. 

Even in the era of AI-assisted research, we believe that research ultimately serves human understanding, and that we will never give up our role in deciding whether a given result is understandable and important. This is going to be a community effort, and as always we value your feedback.