Towards decentralized ML conferences
Published on 2025/07/18. Thanks to Chloé-Agathe Azencott, Linus Bleistein, Edwige Cyffers, and Margaux Zaffran for helpful discussions and comments.
Abstract: Large ML conferences play a central role in our community, by providing a rigorous review process and a venue for meaningful discussion. However, they now face challenges due to their rapid growth, with a number of presented papers and attendees multiplied many times over in the past decade. The massive scale of the largest ML conferences threatens their original purpose of fostering open, community-wide interaction, transforming them instead into reunions for a fortunate few and an overwhelming experience for many others. To address these issues, we propose to revisit the traditional centralized conference model in favor of a more decentralized organization. As a first step, local meetups have been organized shortly before the main conference. However, their limited scale prevents them from fully addressing the shortcomings of large conferences. We are therefore particularly excited by the launch of EurIPS this year— the first NeurIPS-endorsed satellite event held in Europe simultaneously with NeurIPS. This collaboration between NeurIPS and the European ML community is very welcome, in order to better accommodate the diverse needs of researchers and ultimately create a better NeurIPS for everyone.
What is wrong with large centralized conferences?
The current state of large centralized conferences is unsatisfactory, uninclusive, and unsustainable.
Unsatisfactory, because the scale of the event makes it difficult to form new connections or engage meaningfully with the scientific content. For many attendees, especially early-career researchers unfamiliar with the community, navigating a sea of attendees, posters, and talks often feels like searching for a needle in a haystack. This has led to widespread frustration among researchers we've spoken with—many senior scientists have stopped attending altogether, while junior researchers often feel they must endure the experience until they reach a more established position.
Uninclusive, because access to major conferences is limited to a subset of the community. Limited lab funding, disabilities that make travel impossible, caregiving responsibilities, and visa restrictions all turn participation into a logistical and financial hurdle. As a result, conferences have become the meeting point of a small fraction of the research community, putting a considerable part of researchers behind, even more when many of us cannot or will not cross the American border.
Unsustainable, because air travel to conferences is a major contributor to global CO₂ emissions. Reducing these emissions is critical if we are to mitigate the increasing severity and frequency of climate-related disasters. For a more detailed discussion on the environmental impact of conferences, see Appendix A.
All in all, the stakes for the ML research community could not be higher: to preserve space for researchers to meet and exchange ideas, ensure inclusivity for all community members, create opportunities for early-career researchers to join the discussion, and commit to a transition toward more sustainable practices. Otherwise, we believe researcher disengagement from large ML conferences will accelerate, potentially leading to their decline. To avoid this outcome and keep a vibrant and engaging global conference system, we propose a gradual shift from fully centralized conferences to a more decentralized model.
A first step in this direction would be to retain a central venue, while simultaneously setting up regional hubs. Those hubs would each nost no more than 2000 people, to maintain a manageable scale, while collectively accommodating a significant fraction of the conference community. The scientific program at those hubs would be similar to the main venue---featuring presentation of accepted papers through posters and talks, keynotes, workshops, etc. Where time zones allow, keynotes could be broadcast between hubs and the main venue. In the longer term, one could envision phasing out the central venue entirely, relying instead on a network of regional hubs. A key factor for the success of the hubs is their official endorsement by the main conference. This should include financial support through the transfer of registration fees, as well as allowing authors to present at a hub instead of the central venue.
Such an approach would address many of the challenges outlined above. Smaller-scale venues would foster higher-quality scientific discussions and make it easier to connect with new colleagues. Regional hubs would also create more visa options, reducing barriers for those unable to travel to a single central location. Finally, shorter travel distances would reduce logistical burdens and significantly cut carbon emissions.
A proof of concept for a decentralized conference system: NeurIPS in Paris
A first step towards decentralization was actually already introduced during Covid through local meetups. Although their official endorsement ended after the pandemic, these gatherings have continued in some locations. In particular, since 2021, we have been organizing an annual event called NeurIPS in Paris. This two-day meetup takes place the week before NeurIPS and primarily features oral and poster presentations of NeurIPS papers (see Appendix B for more details on the organization of NeurIPS in Paris). The most recent edition brought together around 300 researchers and showcased 100 posters. Attendees came mainly from France and neighboring European countries. Feedback has been overwhelmingly positive, with participants highlighting the quality and depth of discussions made possible by the smaller, more focused format. These experiences strongly suggest a high demand for smaller-scale, high-quality ML research venues.
However, while they provide many benefits, independent local meetups held the week before main conferences do not fully address the issues highlighted above. Indeed, local meetups are orders of magnitude smaller than the main conference, which limits their overall impact and reach within the community. Moreover, under the current model, they lack official endorsement and support from the main conference.
For these reasons, we believe it is essential to take a further step, by setting up hubs at the same time as the central venue and officially endorsed by the conference, including giving the option to present accepted papers at hubs rather than at the central venue.
EurIPS, a new step towards a true decentralized conference system
Establishing a parallel hub alongside the main conference venue is an ambitious endeavor that requires several key ingredients: strong demand from the community, motivated local organizers, local institutional support, and official endorsement from the main conference.
We are thrilled that, for the first time this year, all these elements have come together. NeurIPS is officially endorsing, as an experiment, parallel presentation venues in addition to the main venue in San Diego. There will be one hub in Copenhagen, Denmark (named EurIPS) and one in Mexico. EurIPS will feature presentations of accepted NeurIPS papers, keynotes, and workshops (we encourage you to consider submitting a workshop proposal to EurIPS!).
We are very happy that the NeurIPS board and the European ML community collaborated to make this event happen. It is our strong belief that EurIPS will strengthen the appeal of NeurIPS, and ultimately make attending NeurIPS a better experience for everyone. EurIPS aligns closely with the vision of a decentralized conference system outlined above.
However, there is one important caveat: under the current NeurIPS policy, at least one author of each accepted paper must register for the San Diego venue. This is a departure from the policy in place until 2024, when a virtual registration was sufficient. We find this change both problematic and poorly timed. It is problematic, because many researchers cannot, or prefer not to, travel to San Diego for reasons discussed earlier, and yet this policy change was not clearly communicated to authors before the submission deadline. It is poorly timed because, for the first time this year, a viable alternative is being offered. Mandating San Diego registration under these circumstances is particularly detrimental, as it undermines the potential of these new parallel venues, and ultimately is damageable to the credibility and inclusiveness of NeurIPS.
To truly accommodate the diverse needs of the community, we believe it is essential to allow authors who cannot or do not wish to present in San Diego to instead present their work at other officially endorsed venues. This would ensure that all researchers have a fair and equitable opportunity to share their work at NeurIPS.
Appendix A: Air travel to conferences and sustainability
One week after ICML 2023 in Hawaii, major wildfires broke out, killing over 100 people in the most lethal wildfire in the USA in over a century. Beyond this coincidence, such extreme meteorological events are poised to become more frequent in the future due to climate change. To limit their impact on natural habitats and human lifes, it is critical to curb CO2 emissions in all human activities. Research and in particular ML research are no exception. In our trade, air travel to conferences is a major source of emissions (alongside GPU usage, see below). For example, at the scale of a large research university like EPFL, academic trips across all disciplines represented around 14,000 tons of CO2e in 2019 (pre-pandemic), which corresponds to approximately 3.5 tons per researcher (source 1, source 2). This amounts to around two transatlantic flights per researcher per year, and largely exceeds by itself the individual CO2 yearly target fixed by the Paris agreements (2 tons). As a consequence, limiting academic travel is essential to align our activity with the Paris agreements. In particular, this goes through rethinking the conference system to lessen the carbon impact of conferences.
There are two main counter-arguments often heard when discussing this matter with fellow ML academics:
- “Flights to ML conferences are an important matter, while accounting for little of the global emissions, let’s tackle the rest first.”
The argument is known as whataboutism, that is, deflecting attention to another issue to avoid addressing one’s own. However, to mitigate climate change, emissions in all sectors should go to zero by 2050. There is no exception. Going beyond, as academics, we should arguably be leading---or, at the very least, not be lagging behind---societal transformation in the light of what science tells us of climate change.
- “Emissions caused by GPUs are much larger than the ones of flights, so you should be working more on developing frugal models and less on air travel.”
Actually, the comparison is not that clear, especially for those of us working in academia, who have less access to computational resources compared to our peers in industry. For example, 3.5 tons of CO2e emissions correspond to 5x10^21 FLOPs of compute on recent TPU architectures (660 gCO2e/1e18 FLOPs, cf Table 1 in this paper. This figure does not account for compensations by CFE purchases, which would further reduce the impact of compute.). This is 200x the compute needed to train the original Big Transformer architecture as reported in Vaswani et al. (which is the experiment which convinced the whole community to use Transformer). While arguably we are in a different world now with LLMs, there are also many of us for which this scale already provides ample opportunity for exciting research. Besides, more fundamentally, the argument about GPU emissions is again an instance of whataboutism: it is not because GPUs pollute that we should not care about the emissions due to flying.
Appendix B: More details on NeurIPS in Paris
Since 2021, NeurIPS is Paris has been a two-day event happening the week before NeurIPS, which mainly consists in oral and poster presentations of NeurIPS papers. The event is built with four key principles in mind.
First, the event is completely free of charge, which is made possible by the support of public and private sponsors. We additionally offer travel grants to researchers facing financial difficulties. The downside of running the event free of charge for participants is that this creates large demand, as well as a sizable proportion of no-shows.
Secondly, we keep the event of a reasonable size (around 300 participants), significantly under the number of people who could be interested in the event. This aims at balancing objectives of creating a sense of community in the ML research ecosystem in Paris, as well as keeping logistics under control, while at the same time serving as many attendees as possible. The size of the event makes it manageable by a core team of 3-4 well-motivated researchers on top of their usual duties.
Third, our event is meant to be as inclusive as possible: given the size constraint, we favour young researchers when selecting attendees, and include representation concerns in building the scientific program. Additionally, our conference features several talks on the social aspects of research, including biases, inclusivity and ethics. These talks are fully integrated in the flow of the conference and not isolated during a special session. We also offer mentoring sessions with a diverse panel of researchers that can help students build their career.
Finally, we have designed our event to be as sustainable as possible. We do not encourage or financially support air travel, and have set the location in a central European city which is easy to reach by train from a variety of locations. Our catering is completely vegetarian. Sponsors are asked not to bring any goodies in order to limit waste. We also favour research addressing the social aspects of machine learning when selecting oral presentations.