2026 Conference

This year, the NeurIPS 2026 Position Paper Track made the decision to require that all papers be substantially human-written, with AI used for only copy-editing or similar peripheral changes to the main text. While we recognize that thoughtful use of AI can result in productivity gains in research, the use of AI to write papers creates an acute risk for the peer review system. As Position Paper Track chairs, we took a conservative approach in policy this year as we believe in argumentative work like position papers, excessive use of AI in writing submitted papers has little benefit for the research community as a whole. AI-generated text is often slick, but can depart significantly from the authors’ original intention. In this case, submitting AI-generated text for peer review externalises the cost of verifying that work, imposing it on reviewers. Where AI-generated text is not itself incoherent or otherwise misguided, this raises questions about the appropriate attribution of credit.

To assess if authors were largely abiding by this policy, we partnered with Pangram, a leading AI detection modeling company. We worked closely with Pangram to ensure, as per their enterprise-level data agreement, that zero data would be retained through the usage of their model. After several independent analyses to verify the correctness of this model and rule out scenarios in which significant false positives would be created, we are now making the difficult decision to uphold our policy, under which:

• 178 submissions (18.4% of all submissions) will be desk rejected
• 123 submissions (12.7%) will be requested to provide evidence of substantial human engagement or risk a desk reject. 

In this blog post, we will lay out our analyses informing this decision, and provide our perspective as organizers.

Why this policy?

We reproduce here the 2026 PPT AI policy:

Use of AI: While we recognise the productivity gains that can be realised through judicious use of AI in research, due to the risk to the integrity of individual projects and of the review system as a whole, the position paper track is establishing the following explicit guardrails on AI use in preparing and reviewing submissions.

• While AI tools may be used in the research that leads to the final paper, the final paper must itself be substantially written by human authors, meaning that AI is used only for copy-editing or similar peripheral changes to the main text.
• At submission time, authors will be required to state how AI tools were used in the preparation of the paper, if at all, and to attest that they have not used AI in ways contrary to the above rule.
• Because papers submitted to the position paper track are confidential, reviewers will be required to commit to not using AI tools to write their reviews.
• Reviewers and authors found to have contravened their commitments not to use AI may be subject to desk-rejection of any work submitted to the position paper track. 
• Note that the Position Paper Track’s LLM policy differs from the Main Program’s LLM policy. Authors are responsible for understanding policy pertaining to the specific track they are submitting to, and abiding by it.

The use of AI to write papers creates an acute risk to the peer review system. Proactive steps are necessary to build the norms and institutions that will preserve its integrity. This policy is an attempt to begin that process. 

It is of course possible that a paper’s authors could use AI responsibly, (1) personally verifying every line of AI output, and (2) ensuring that the AI does nothing more than rephrase ideas for which humans are solely responsible. However, by submitting work that is immediately recognisable and verifiable as being substantially AI-generated, authors make it impossible for readers to know that (1) and (2) obtain, leaving reviewers with little choice but to rely upon author declarations. Unfortunately, given the volume of submissions that appear non-compliant, relying on author declarations is insufficient.

We do not expect that our policy and our approach will be the last word on handling AI-generated research. Every research field will have to confront the same problem, and a range of solutions may be reasonable. We have sought to use the evidence available to us to identify submissions that appear to be non-compliant with our policy. But we are also introducing a new approach to auditing AI use by establishing appropriate provenance.  Authors whose submissions show significant AI involvement must provide an audit trail that clearly demonstrates that they complied with the policy. We expect that in future years this kind of audit trail will become a default. 

AI detection with Pangram suggests substantial AI use among this year’s submissions

We identified if a submission is significantly AI-written using Pangram, an industry-leading AI detector. Using Pangram (v3.3.2), we found that 28.2% (273 / 969) of submissions substantially used AI for writing. This finding prompted further investigation, which we present in the next sections. We start by providing clarity on what Pangram does.

Given a full text document, Pangram first uses a windowing algorithm to break up the text into text windows, where by default, each window is around 250 to 350 words. Next, Pangram assigns each text window a probability that it contains AI-generated text. If the model’s assigned probability exceeds 0.75, then that window is flagged as AI-generated. From these predictions, each paper receives a Pangram AI score, which is the percentage of windows that are classified as AI-generated. A Pangram AI score of 100% means that all of the words in the paper fall into a text window that Pangram believes contains AI-generated text. A Pangram AI score of 100% should not be interpreted as “100% of the text is AI-generated”, rather that there is substantive use of AI in many parts of the text.

Our preliminary investigation found that 28.2% (273 / 969) of submissions to the NeurIPS 2026 Position Papers Track (PPT) received a Pangram AI score of 100%. We found this number surprisingly high, given internal and external audits of Pangram reported a false positive rate of less than 0.1%, and in previous applications to ICLR 2026 accepted papers, the model only detected that 1% of papers were AI-generated. We contrasted Pangram’s results on the NeurIPS PPT against papers from comparable venues (Table 1). We tested Pangram against papers accepted to ACM FAccT in 2022 and 2025, which are similar in style and content to many NeurIPS position papers. FAccT 2022 papers preceded ChatGPT’s release and served as a negative control. To determine if our findings extend to other NeurIPS tracks, we compare against a sample of 2025 and 2026 submissions to the NeurIPS Evaluations and Datasets (E&D), formerly Datasets & Benchmarks (D&B).

Table 1
Default Pangram AI-detection across conferences.

Conference	# Papers	Pangram AI Score
		≥ 50%	≥ 90%	= 100%
NeurIPS PPT 2025	536	28.5%	11.9%	8.2%
NeurIPS PPT 2026	971	70.5%	42.7%	28.2%
NeurIPS D&B 2025	996	5.6%	0.8%	0.4%
NeurIPS E&D 2026	996	43.7%	9.3%	2.1%
FAccT 2022	159	0.0%	0.0%	0.0%
FAccT 2025	204	1.0%	1.0%	0.0%

We made two observations. First, there are far fewer papers with a Pangram AI score of 90-100% in NeurIPS E&D and FAccT compared to the NeurIPS Position Paper Track. Second, there is a sharp increase in AI use for paper writing in both NeurIPS tracks evaluated; in the Evaluations and Datasets track, papers with a Pangram AI score ≥90% have increased more than tenfold from 2025 to 2026. Taken together, this suggests the high rate of AI use in the NeurIPS Position Paper Track is caused both by factors specific to the track itself, and by a broader significant increase in AI use across the board. 

Using smaller text windows leads to more localized AI use at the cost of recall

One challenge to our preliminary findings that “28.2% of submissions have 100% Pangram AI scores” is that Pangram classifies on large text windows (250-350 words, by default), and it is possible that Pangram flags a text as AI-generated, even though only a small portion of the text was written by AI while remaining compliant with our policy. We re-run Pangram using two custom text windowing strategies with strictly fewer words: medium-sized (approx. 100 words) and small-sized (approx. 50 words).

Using smaller window sizes reduces the chances of over-claiming AI use, but it may also worsen the ability of Pangram to truly identify AI-generated text. We assess how window size affects recall on 10 ChatGPT-generated “position papers” (Table 2).

Table 2
Comparison of smaller text windowing strategies and thresholds on Pangram AI score.

Papers	Windowing	Avg. Pangram AI score	Recall at ≥ Pangram AI score
			≥ 0.5	≥ 0.7	≥ 0.9	= 1.0
ai_positions25(N=10)	small	61.8%	70%	30%	0%	0%
	medium	91%	100%	100%	70%	0%
	default	100%	100%	100%	100%	100%

These results suggest that 100-word windows result in a lesser drop in recall compared to 50-word windows, so we decided to move forward with medium-sized windows, trading off recall for finer-grained claims on AI use. Using medium-sized windows, the percentage of papers with Pangram AI scores of 90-100% goes down from 42.7% to 12.7% (Table 3).

Table 3
Varying window size on Pangram AI scores in NeurIPS PPT 2026.

Window size	Pangram AI Score
	≥ 50%	≥ 90%	= 100%
medium	62.3%	12.7%	2.16%
default	70.5%	42.7%	28.2%

To ground our findings, we tested Pangram on several writing scenarios with varying AI involvement. We selected 10 papers from FAccT 2022 that resembled position paper track submissions. For each, we extracted a random 100 word text window. Using OpenAI’s GPT 5.5 via OpenRouter, we tested 12 AI use cases. In Table 4, we categorize each use case by their permissibility against our stated policy. We performed two additional experiments. We tested Pangram’s sensitivity to obvious LLM instruction-following text (e.g., “Sure, here is your paragraph”), which we term “AI residue”. Lastly, we tested how sensitive Pangram is to increasing percentages of AI-generated text; we do this by truncating the original text at different amounts from 5% to 95% and asking the LLM to complete the remaining text.

Table 4
AI use cases and permissibility.

Breaks policy?	Use Case	What it tests
Clearly permissible	Proofreading	Request that an LLM edits only spelling, punctuation, grammar, and citation-format cleanup.
	Light copyediting	Request that an LLM edits only local clarity, concision, awkward phrasing, and sentence-level polish, with no substantive change.
Borderline permissible	Heavy copyediting / line editing	Request that an LLM edits large wording changes and sentence restructuring, while preserving the same claims and reasoning.
	Structural rewriting	Request that an LLM reorganizes paragraph or argument presentation while preserving the human’s ideas.
	Hybrid revision	Human and AI both materially shape the prose, including back-and-forth assistant use or human paraphrasing after AI edits. Tested with Codex, and 5 editing turns (original, AI edits, human edits, AI edits, human edits).
	Translation / backtranslation	Request that an LLM translates between languages, so that meaning is preserved, but surface wording may be extensively replaced.
Clearly impermissible	Generation from a single-sentence human plan	A human writes a one-sentence plan/thesis, then AI generates the full passage from it.
	Substantive AI rewriting	Request that an LLM changes claims, reasoning, framing, or argumentative structure.
	Original AI-authored passage	Request that an LLM writes a new position-paper-like passage from examples, topic, or instructions.
	Human edits AI work	Human makes minor edits to an original AI-authored passage.
Diagnostic tests	AI residue	Insert obvious chatbot artifacts or AI-styled residue into otherwise human text (e.g. “sure, here is your paragraph:)
	Partial AI completion	AI receives part of the original human text and completes the rest. Conditions: AI completes 5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 95%

For all of the permissible uses, Pangram did not classify any as AI-generated, despite a substantial change in the text (Figure 1A). Meanwhile, clearly impermissible use cases were flagged by Pangram as AI-generated. In our experiment on partial AI completions, Pangram never classified AI completions of 20% or less as AI-generated (Figure 1B). While these experiments were conducted with only 10 text samples, the findings suggest that papers with very high Pangram AI scores were not compliant with our AI use policy. 

Figure 1

Experiments on AI use and partial AI completion.

Actions to uphold our AI use policy

Our policy stated that AI may be used only for “copy-editing or similar peripheral changes to the main text”. In recognition of the fact that norms in this area are emerging, and these terms are inherently ambiguous, we have adopted conservative decision thresholds designed to minimize false positives (see Table 5 below). We have also incorporated additional supporting evidence, such as authors’ AI use declarations, evidence of policy non-complicance in other submissions, and authors whose submission patterns – including a high number of solo-authored papers – warranted additional scrutiny.

For submissions found to be very likely in breach of our policy, we have chosen two courses of action, based on the strength of the evidence that they are in breach.

• Where we found sufficient evidence of non-compliance with our AI policy, we have issued a standard desk-rejection that is not subject to appeal under standard circumstances.
• Where we have strong but not decisive evidence of non-compliance, we are asking authors to provide evidence supporting that their AI use complied with policy, with details below. Submissions lacking such proof by June 15th, 2026 will also be desk-rejected. 

Table 5
Decision Thresholds.

Outcome	Pangram AI Score	Additional Considerations	#
Desk Reject without Appeal	≥0.9	None	77
	≥0.8	Author has submitted multiple solo-authored papers, with at least one above this threshold OR At least one author has at least one other desk reject	79
	≥0.5	Author declared that they did not use AI, or did not declare AI use	22
Total (% of total submissions)	178 (18.4%)
Desk Reject with Appeal	≥0.8<0.9	None	123
Total (% of total submissions)	123 (12.7%)

Appeals Process

It is important to be clear where the burden of proof should lie when it comes to detecting improper use of AI in submissions to a peer-reviewed conference. 

Authors who use AI extensively but, in their belief, responsibly, are often indistinguishable from those who have used AI in ways that are not consistent with policy. We have used every available measure to distinguish between these two groups, but we acknowledge there will inevitably be borderline cases. 

We believe that it is inappropriate for the research community as a whole to bear the cost of making more fine-grained distinctions among these cases. Authors who limit AI use in their final drafts help reduce burden on the review community. Authors who make more extensive use of AI should maintain clear documentation of their process, which can be shared if requested.

Authors given the opportunity to appeal will be able to provide evidence of responsible AI use in the following form. Well-motivated equivalents may be considered. 

• Authors must supply the Track Chairs with a link to an online version of their paper that has a version history including the work before and after the use of AI
• They must identify (1) a “pre-AI” checkpoint indicating that they developed the substantive content of the paper independent of AI, (2) a “post-AI” checkpoint immediately after their most substantive AI-written edits, and (3) the final paper as submitted.
• They must present analysis showing that the AI edits in (2) did not introduce new substantive content that was not present in (1), and of human edits after (2) which demonstrate that (3) was appropriately verified by human  authors. 

This dossier will be reviewed by the PPT team. Authors who do not wish to appeal may withdraw their paper. We recognize this is a difficult and consequential situation for affected authors, and we encourage anyone with questions to reach out.

Reflections and perspective

The research community faces two overlapping challenges. The first is the impact of irresponsible AI use on our peer review institutions. The second is building consensus on AI use policy, acknowledging there may not be a solution that works across every publication venue, as the community is in substantial disagreement about how to deal with the first. 

In formulating our policy on AI use, we explicitly recognised that the careful use of AI can benefit research. We focused narrowly on prohibiting excessive AI use in writing the submitted paper, for the reasons detailed above. We attempted to capture what we believe to be both a reasonable and widely-supported approach to handling AI use, but the very fact that so many authors will face adverse consequences entails that there is complexity in this matter.

Building consensus around the appropriate use of AI will require the development of new norms. The norm we have attempted to implement here is one that holds that the substantial use of AI for writing final submissions imposes a significant cost on the research community: it becomes unclear if an AI-generated text includes only human-generated and human-verified ideas, for which humans will take full responsibility. We believe that in order to justify these costs on an already-strained system, there must be a corresponding benefit to the community. We hold that delegating the writing of the final paper to AI does not offer that benefit. 

We thank the community for their sustained interest and engagement in the NeurIPS Position Paper track. We hope these decisions will not only ensure that authors are abiding by stated policy, but also elevate the quality of submissions that reviewers and area chairs will be dedicating time to this year. 

Alex Lu, Seth Lazar, David Rugamer
NeurIPS Position Paper Chairs

Stanley Hua, Kate Metcalf
NeurIPS Assistant Position Paper Chairs

Ethics in ML research is an important part of ensuring the integrity and impact of scientific work, and NeurIPS 2026 is looking for Ethics Reviewers to support this process. If you are able and willing to participate in the review process, please sign up at this form. Feel free to share this call with your colleagues.

Key Dates

We ask that ethics reviewers

• Review up to 5 papers each,
• Provide ethics reviews during at least one of the following periods:
  • July 6 – 20, 2026 (main ethics review period), * 
  • July 22 – August 13, 2026 (emergency review period – review requests may arrive at any point during this period)

* Note that ICML 2026 takes place July 6-11.

A full list of relevant dates for the conference is available here.

Again, the sign-up form is here: NeurIPS 2026 Ethics Reviewer Self-Nomination 

About the ethics review process

The main reviews conducted through the program committee (reviewers, program chairs, and area chairs) are, and continue to be, the sole decision-making process for accepting or rejecting papers for publications at NeurIPS. Reviewers are expected to review submissions not just for pure technical merit, but also in the context of the NeurIPS Code of Ethics.

The ethics review is a second round of review that takes place mainly when the program committee flags any potential concerns during the main review phase that merit further attention. Ethics reviewers provide feedback to the program committee regarding risks and harms of the work in line with the NeurIPS Code of Ethics, and recommend potential mitigations for authors to incorporate when revising their submissions.

The ethics review process is not a disciplinary or punitive process. However, in rare situations, the NeurIPS program committee may decide to reject submissions that have grossly violated the NeurIPS Code of Ethics, taking into account recommendations from the ethics reviews. In past instances where this occurred, the authors were provided with substantial guidance and relevant citations and were invited to revise and resubmit to NeurIPS.

As detailed in the Ethics Guidelines for Reviewers, the ethics reviews generally follow the double-blinded review process of the main reviews. ​However, additional steps are taken in order to minimize exposure risks. During the ethics review process, any submissions flagged for ethics review will not be publicly labeled as such. During the author response period, ethics reviews will be anonymized when made visible to authors and main reviewers. When the final accept or reject decision has been made, authors who have accepted papers may, at their discretion, choose to make their ethics review public.

Thank you for your consideration,2026 ETHICS REVIEW CHAIRS
Stephanie Hyland, Principal Researcher, Microsoft Research
Emanuel Moss, Senior Research Scientist, Intel
ethics-review-chairs@neurips.cc

When Joelle Pineau launched the first Machine Learning Reproducibility Challenge at ICLR 2018, it was a small community experiment: could we systematically invite researchers to reproduce published results and share what they found? 

Eight years and eight editions later, we are delighted to announce that MLRC 2026 will be an official track at NeurIPS 2026 –  the first time in MLRC’s history that reproducibility science has a dedicated home inside a major ML conference.

This milestone reflects the community’s growing relationship with reproducibility. NeurIPS wanted to send a meaningful signal to the field that reproducibility has become a scientific question worthy of its own rigorous study – this official track allows that to happen.

From challenge to venue

The early iterations of MLRC (v1, v2, v3) were structured as challenges: pick a paper, try to reproduce it, and report what happened. They were enormously valuable as an educational exercise, especially for early-career researchers. Courses like FACT AI at the University of Amsterdam built entire curricula around the challenge, and the quality of the work that came out of those courses showed just how seriously students took the opportunity.

As we reflected on what we learned across those iterations, it became clear that MLRC should be more. Reproducibility is not a binary outcome: a paper is never simply “reproducible” or “not.” The most insightful submissions were always the ones that engaged deeply with the specific claims of a paper, pushed those claims into new settings (generalization), tested their limits, discovered novel insights on top of the paper’s claims, and reported back with nuance. We wanted to create a venue that actively sought out that kind of contribution. Over the years, we improved the program to incentivize these submissions. 

Reproducibility studies are also not easily rewarded by the ML community’s standard metric of novelty: they do not propose new architectures, beat state-of-the-art numbers, or introduce new datasets. Getting researchers to invest serious effort in this kind of work required building a publication and recognition path that made that investment worthwhile. We have updated this systematically over the years: early iterations published through ReScience, a respected open journal for reproducibility across computational science; then, in 2023, we transitioned to TMLR, bringing MLRC papers into a high-prestige, well-indexed ML venue with a rigorous open review process. 

While the initial MLRC operated primarily as satellite workshops to conferences, in 2022 (and 2023), we partnered with NeurIPS in the Journal to Conference track, where the reproducibility papers were presented in poster sessions alongside the conference papers. In 2025, we ran MLRC’s first in-person conference at Princeton University, further elevating the incentive in the form of a dedicated conference and physical stage, inviting keynote speakers to present, and having a full-day event with orals, posters, and networking sessions. Each step raised the incentives, and the NeurIPS track is the next step in that progression.

What this means in practice

MLRC 2026 accepted reproducibility papers will be presented in person at NeurIPS 2026 in Sydney, Australia (December 6–13, 2026), alongside papers from the Main Track and the Evaluations & Datasets Track. The submission and review process remains anchored in TMLR. Reproducibility papers must first be accepted at TMLR within the eligibility window, and then undergo a light compatibility review by the MLRC committee to confirm suitability for the track.

This TMLR-first model is deliberate. TMLR’s open, continuous reviewing cycle allows authors to refine their work and get expert feedback before it is considered for presentation. It also means that accepted papers carry the full weight of TMLR’s review standards, independent of MLRC. The MLRC committee’s role is then to identify papers among accepted TMLR submissions that represent the best of reproducibility science and would benefit from the visibility of a NeurIPS venue.

What we are looking for

MLRC has always welcomed a broad range of reproducibility work, and that continues this year. We are looking for papers that take reproducibility seriously as a scientific question — not just as a means to an end, but as a contribution in its own right. This includes:

• Reproductions and replications that rigorously test specific claims from published papers, whether they confirm, partially replicate, or fail to reproduce prior results
• Generalizability studies that extend original findings to new settings, datasets, or model architectures, adding insights that the original paper could not offer
• Meta-reproducibility studies examining reproducibility patterns across a body of related work
• Methods and tools that make reproducibility research more accessible or rigorous
• AI-assisted reproducibility, including studies that use or critically evaluate automated approaches to replicating research papers
• Reproducibility of AI systems and agents as subjects of study in their own right

An important point to note: negative results and partial failures to reproduce are as valuable as confirmations. Science advances by understanding where claims hold and where they do not. A careful, well-documented failure to reproduce a result — with a clear account of what was tried and what was found — is a genuine contribution to the literature.

Note: Work focused on evaluation methodology more broadly may also be a good fit for the Evaluations & Datasets NeurIPS 2026 track. We encourage authors to consider both venues when deciding where to submit.

How to submit

To be eligible for MLRC 2026, your reproducibility paper must be accepted as is / with minor revisions to TMLR and must have been submitted to TMLR between June 20, 2025 23:59 AOE and September 30, 2026, 23:59 AOE. Please check our CFP for more details.

We accept submissions through three paths:

Path 1: Expression of interest (EOI) before acceptance. If your reproducibility paper is currently under review at TMLR and was submitted within the eligibility window, you may submit this EOI form to express your intent to be considered for MLRC. The “intent to submit” deadline is June 4, 2026, AOE – this is primarily a soft deadline to allow you to submit your paper to TMLR well in advance, so that you get your decisions in time. If your reproducibility paper is subsequently accepted to TMLR, you will be asked to update the form with your acceptance details and camera-ready materials. The hard deadline by which we need your TMLR paper’s acceptance decision is September 30, 2026, AOE – we cannot accommodate your paper into MLRC any later than this date, even if you are still waiting for your TMLR decisions.

Path 2: Self-nomination after acceptance. If your reproducibility paper has already been accepted to TMLR and was submitted within the eligibility window (start date June 20, 2025 23:59 AOE), you may submit this form to be considered for MLRC. Once accepted to MLRC, you will be asked to update the form with your camera-ready details. Similar to Path 1, the hard deadline to submit this form is September 30, 2026, AOE. Ensure you have not submitted your accepted paper to the NeurIPS 2026 Journal to Conference track, as our dual submission policy restricts dual presentation.

Path 3: Area Chair nomination. TMLR Area Chairs may nominate accepted reproducibility papers within the submission window for consideration at MLRC. No action is required from authors unless they are contacted. Area Chairs can use the same form to submit their nomination. The deadline for TMLR Area Chairs to nominate reproducibility papers is also September 30, 2026, AOE.

For all paths, please read the TMLR author guidelines and submission guidelines for formatting and review process details.

The hard deadline to have your TMLR acceptance in our system is September 30, 2026 AOE.

Important Dates

• Earliest date of the acceptance of your TMLR reproducibility paper for consideration to this year’s MLRC: June 20, 2025 23:59 AOE
• Soft deadline (EOI / intent to submit): June 4, 2026, 23:59 AOE
• Deadline to have TMLR decisions in our system: September 30, 2026, 23:59 AOE
• Author notifications: October 7, 2026
• NeurIPS 2026: December 6–13, 2026

Closing Thoughts

Reproducibility has always been foundational to science. What MLRC has tried to do, across eight editions, is make reproducibility a first-class research activity in machine learning — one that is worth investing in, publishing, and being recognized for. Having MLRC as an official NeurIPS track is an affirmation that the community values this work, and we hope it encourages more researchers to take reproducibility seriously as a scientific contribution.

We look forward to seeing the community’s work at NeurIPS in Sydney. Please visit the MLRC 2026 website for full details, and do not hesitate to reach out at  at reproducibility-chairs@neurips.cc with any questions.

This year, to improve transparency and responsible use of datasets, we are introducing a new requirement for dataset submissions in the NeurIPS 2026 Evaluations and Datasets Track: all dataset submissions must now include Responsible AI (RAI) metadata as part of the dataset’s Croissant file.

Dataset submissions to the track are already required to include a Croissant file, a standardized, machine-readable metadata format for describing datasets and their structure. We extend this requirement to also include RAI metadata within that file.

RAI metadata helps ensure datasets are used appropriately in AI research. Providing standardized information on how a dataset was created, its limitations, and its intended use helps researchers understand when and how the data can be used, reducing the risk of misuse, biased results, or misleading conclusions. This requirement also advances a core goal of the track: making datasets more transparent, comparable, and reusable across the community.

To make this as smooth as possible, we collaborated with the Croissant RAI team to establish a minimal set of RAI fields describing key aspects such as dataset limitations, potential biases, intended use, and other responsible AI considerations. Authors must include the RAI information directly within the Croissant metadata wherever possible, or clearly link to relevant sections of the paper when needed.

To support authors, we provide two tools to streamline this process:

• An online RAI editor to help complete the required RAI metadata fields
• A Croissant validator to ensure files are complete and correctly formatted

We recommend completing these steps as early as possible, and inform authors that dataset submissions missing RAI metadata from their Croissant file will be flagged during review.

For full details on the dataset hosting and submission requirements, please see the Call for Papers and the data hosting guidelines. 

Joaquin Vanschoren, Konstantina Palla, Jessica Schrouff, Alexandre Drouin, Lijun Wu

Evaluations and Datasets Track Chairs 2026

NeurIPS Program Chairs: Marc Deisenroth, Finale Doshi-Velez, Nika Haghtalab, David Rolnick, Jenna Wiens

Google Research: Rajesh Jayaram, Vincent Cohen-Addad, Drew Tyler, David Woodruff

Google Cloud: Jinsung Yoon, Mihir Parmar, Palash Goyal

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

NeurIPS Communication Chair: Jean Kossaifi

Following positive feedback from other venues, like STOC and ICML, NeurIPS is pleased to announce a new initiative, in partnership with Google, that will provide authors access to their Paper Assistant Tool (PAT) and support authors in improving their submissions to NeurIPS 2026. 

This program offers authors a limited opportunity to receive free, automated, and actionable feedback on their manuscripts before the final deadline, private to the authors. The feedback the authors will receive from PAT through the NeurIPS program will not be used in the review process. Reviewers, area chairs, and program committee members will not have access to the PAT feedback. 

What is PAT?

PAT is a specialized, experimental tool powered by Google’s Gemini models, utilizing a “reasoning”-focused pipeline. 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, (see blog post), 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. 

While the program at STOC focused heavily on theoretical correctness, the ICML program was then specifically tuned to address the needs of the machine learning community. After considering the feedback received during the ICML program, including pain points raised by authors, the Google Research team augmented PAT by integrating components of the ScholarPeer system into the PAT feedback.  As a result of this integration, compared to the ICML iteration, this version of PAT has:

• Improved search and tool capabilities: allowing for fact checking and reducing the hallucination rate.
• Improved comparison with related work: resulting from the ScholarPeer deep literature review agent. This empowers PAT to take into account related work in its analysis of the paper.
• Strengths / Weaknesses Analysis: the feedback produced by PAT will contain an analysis of the potential strengths and weaknesses of the paper. This will enable authors to improve their paper at a higher level, rather than just addressing individual errors.

Like the ICML and STOC programs, the goal is to help authors 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, ending on the Abstract Deadline (May 4th, 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. In alignment with the system ultimately used during ICML 2026, we define an eligible author as any author whose OpenReview Account was created prior to April 1, 2026.

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 NeurIPS review process. It is visible only to the authors. 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 NeurIPS review process, including the NeurIPS 2026 AI Assisted Reviewing Experiment. 
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: No one has access to this data unless there is a technical difficulty, in which case Google staff (Rajesh Jayaram and Drew Tyler from the Google Research Organizing Committee) will only inspect the data (submission PDFs and generated feedback) with 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 a human review.

• The model may occasionally flag correct statements as errors or miss actual flaws. It is the author’s responsibility to verify the validity of the feedback.
• Note that the model may make suggestions for the paper that you disagree with. This is not necessarily a bug, and may even be desirable as such behavior could also occur with reviewers. By considering why you disagree with the suggestion, you may be able to add justification to the paper which preempts certain reviewers’ comments, thereby increasing the strength of the paper.

Outcomes

Our primary objective is to empower authors to elevate the quality of their submissions by acting as a high-precision filter. This tool is designed to help authors catch nuanced errors—in proofs, experimental setups, or reasoning—that human reviewers might miss or lack the time to detail.

We believe this tool represents a promising opportunity to test the use of AI to elevate the standard of our own scientific submissions. 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 NeurIPS 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 

The NeurIPS community benefits from a wide diversity of papers and ideas, which can arise and be developed in many different ways. To help cultivate the diversity of papers that belong at NeurIPS, this year the Main Track is asking authors to select a Contribution Type:

• General: We expect that most submissions will fall into this type.
• Theory: The main contribution is via theoretical analyses and proofs.
• Use-Inspired: The main contribution is in framing or designing approaches to meet the needs of a specific real-world application. (This often involves, e.g., engaging with domain experts.).
• Concept & Feasibility: The main contribution is a highly novel, high potential reward idea with scope beyond what can be validated in a single paper. (The significance and originality bar for these contributions is high.)
• Negative Results: The main contribution is in understanding a negative result. (The significance and originality bar for these contributions is high.)

The review form used is the same across all Contribution Types, but the way that reviewing criteria are interpreted will differ across Types. We encourage authors to look at the reviewing guidelines for each Contribution Type in deciding which to select. (By contrast, the Evaluations & Datasets Track and Position Paper Track have different review forms from the Main Track.)

Note that because reviewers will base their recommendations on the Contribution Type, it is not possible for this to be changed after submission, either by the authors or by the reviewers / PCs. It is also, of course, not permitted to submit the same or highly similar papers to multiple Contribution Types or Tracks – such papers will be desk rejected.

To help authors select the right Track and Contribution Type for their paper, we include a number of example papers below.

Example papers

Main Track – General

Segment Anything (Kirillov et al. 2023)

This paper introduces a model (SAM) for segmenting images, and a dataset used for training it. The focus of the paper is on the particular model being developed, so the paper would be a better fit for the Main Track (General) than the Evaluations & Datasets Track. The paper does not motivate or evaluate SAM with respect to the specific needs of a real-world use case – therefore the Use-Inspired contribution type would not be a good match.

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (Frankle and Carbin 2019)

This paper empirically investigates the efficacy of training pruned subnetworks of a neural network. While this area of work is related to deep learning theory, the contributions are not theoretical and so Main Track (General) would be the best choice.

Main Track – Theory

Neural Tangent Kernel: Convergence and Generalization in Neural Networks (Jacot et al. 2018)

This paper formally connects neural network training to kernel methods with the introduction of the Neural Tangent Kernel (NTK). While the authors do conduct some illustrative experiments, the focus is on proving theoretical properties of the NTK and analyzing the implications of these results.

A Universal Law of Robustness via Isoperimetry (Bubeck and Sellke, 2021)

This paper is motivated by the significant overparameterization of deep learning models in comparison to the number of datapoints being fit. The authors prove theoretical results on the number of parameters a model needs to smoothly interpolate between datapoints, showing this is significantly more than the number required for interpolation alone, and consider how these results may align with behavior observed in practice.

Main Track – Use-Inspired

Estimating Canopy Height at Scale (Pauls et al. 2024)

This paper considers the application of ML models to the problem of canopy height estimation. While the core models are not novel, the authors contribute an analysis of what makes this problem and data special and introduce a novel loss function to account for these properties.

Probabilistic Emulation of a Global Climate Model with Spherical DYffusion (Rühling Cachay et al. 2024)

This paper considers the problem of emulating the Earth’s climate over long time horizons. The authors introduce Spherical DYffusion, an approach combining diffusion models with Spherical Fourier Neural Operators, in order to address several challenges particular to atmospheric data. Via extensive domain-informed experiments on a single, well-chosen dataset, the authors compare Spherical DYffusion to other models, both from machine learning and from physics.

Main Track – Concept & Feasibility

Capsule networks (Sabour et al., 2017)

This paper introduces a novel type of neural network architecture, capsule networks, which permit dynamic routing of information via a type of attention computed between different capsule units in the network. The paper shows that capsule networks have promising performance on a range of MNIST-based tasks, but explicitly does not attempt to explore the full range of properties they demonstrate on different tasks and datasets, nor to consider all ways that capsule networks may be instantiated. The authors draw parallels between the early stage of capsule network research and that of recurrent networks several years previously – emphasizing that such promising results in early-stage research suggest the value of more work in this area.

Stochastic Variational Inference (Hoffman et al. 2013)

This paper introduces a scalable approximate algorithm for posterior estimation that allows for analyzing large datasets through a Bayesian lens. This is an example of a paper with very broad applicability. The experiments could obviously not test on all Bayesian models, but provided concrete evidence that the ability to approximately analyze larger datasets provides performance benefits over using traditional inference on smaller datasets.

Main Track – Negative Results

Inherent Trade-Offs in the Fair Determination of Risk Scores (Kleinberg et al. 2017)

This paper investigates the impossibility of satisfying multiple fairness criteria simultaneously. While it relies heavily on a theoretical framing, its main contribution is that of a negative, rigorously demonstrated and surprising result that is not demonstrated via empirical evaluations.

Understanding deep learning requires rethinking generalization (Zhang et al., 2016) 

At the time this paper was written, the common wisdom was that generalization in deep learning was similar to generalization in classical models. Through a combination of experimentation with carefully chosen semi-synthetic data and a theoretical analysis of a specific neural architecture, the authors make a compelling argument that the sources of generalization in classical models – limited expressivity and explicit regularization – cannot explain the generalization properties of neural networks. This was their core (negative) result. While they speculate that implicit regularization via SGD may be a critical element, the negative result is the main argument of the paper.

Evaluations & Datasets Track

ImageNet: A large-scale hierarchical image database (Deng et al., 2009)

This paper offers a dataset for computer vision applications with a demonstration of its value in three tasks. The primary contribution relates to the dataset.

Fairness Through Awareness (Dwork et al., 2011)

This paper gives a definition and implementation of individual fairness, with a secondary contribution that provides an algorithm to improve on this metric. The paper relies heavily on a theoretical framework. While the authors could consider different options (main/general due to algorithm development or main/theory), we believe the Evaluations & Datasets Track is most appropriate given that the main contribution is about the definition of a new fairness metric.

The Illusion of Readiness in Health AI (Gu et al., 2025)

This paper focuses on evaluations for a use-case inspired application in healthcare providing negative results through experimentation. While this is a use-case inspired, negative result, the primary focus of this paper is on experimental evaluations.

Position Paper Track

Position: Bayesian Deep Learning is Needed in the Age of Large-Scale AI (Papamarkou et al., 2024)

This paper revisits the role of Bayesian deep learning in the modern AI landscape. It argues that uncertainty-aware approaches should play a central role. While it is close to a review paper due to its broad synthesis of existing methods, it advances a clear normative stance on the field’s future direction rather than merely summarizing prior work.

Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks (Bechler-Speicher et al., 2025)

This paper examines the current state of graph learning, arguing that progress in the field is hindered by flawed benchmark datasets and evaluation practices. The position is about benchmarking and evaluation practices in graph learning, arguing that they must be fundamentally rethought. While being close to an Evaluations and Datasets paper due to the introduction and analysis of concrete benchmarking setups, it focuses on a broader critique and agenda for the field, which is characteristic of a Position Paper.

Position: Probabilistic Modelling is Sufficient for Causal Inference (Mlodozeniec et al., 2025)

This paper challenges the common view that causal inference requires specialized frameworks such as structural causal models or do-calculus. The position is about the foundations and methodology of causal inference. While lying close to a Main Track paper, it does not introduce a concrete algorithmic pipeline as its primary contribution, but instead advances a conceptual stance on how causal inference should be framed.

We are asking community members to self-nominate as Area Chair for the 2026 Position Paper Track! In line with the track’s aim of expanding topical scope, we are looking for experts from a diverse range of disciplines. NeurIPS relies upon the active participation of the community to evaluate submissions and uphold scientific quality. Area chairs in particular help manage reviewers for submissions, making sure that their reviews are high-quality and on-time, and provide accept/reject recommendations based upon reviewer assessment and discussion. 

Criteria

We welcome self-nominations from anyone who is or wants to be part of the NeurIPS community, and believes they have the experience to act as an area chair. Although we provide some guidance below, especially because we are soliciting self-nominations from a wide range of backgrounds, we recognize that many nominees may qualify as strong area chairs without meeting all (or many) of the criteria below. Thus, we also provide a free-form text box for nominees to justify their readiness. 

To provide some guidelines, we generally expect area chairs to:

• Be relatively senior, meaning they have a mature published research portfolio and experience in a supervisory capacity in projects.
  • A rule-of-thumb is 3-5 last-authored peer-reviewed publications, and 8-10 peer-reviewed publications in any capacity
• To have at minimum two years of experience either as a reviewer or area chair for a major machine learning or related conference (e.g. NeurIPS, ICLR, ICML, CVPR, ICCV/ECCV, ACL, EMNLP, KDD, COLT, FAccT, ACM)

To self-nominate, please fill out our Google forms here. We only accept self-nominations, but we would appreciate anyone who would forward this form to colleagues who would be well-suited: https://docs.google.com/forms/d/1IQWm78nsVFJW7UXU7QgmKXtA2JVSipnSbbCXy609rvU

Alex Lu, David Rügamer, and Seth Lazar
Position Paper Chairs 2026