CtrlK
NeurIPS (the Conference on Neural Information Processing Systems) is one of the most prestigious and influential academic conferences in artificial intelligence and machine learning. Held annually in December, NeurIPS covers topics spanning machine learning, computational neuroscience, statistics, optimization, computer vision, natural language processing, and reinforcement learning. Together with ICML, it is widely considered one of the top two machine learning conferences in the world [1].
The conference was originally known as NIPS (Neural Information Processing Systems) from its founding in 1987 until 2018, when it was renamed to NeurIPS following community concerns about the original acronym.
The conference traces its origins to 1986, when the idea for an annual meeting on neural information processing was proposed at the Snowbird Meeting on Neural Networks for Computing, organized by the California Institute of Technology and Bell Laboratories. The first NIPS conference was held in 1987 in Denver, Colorado, sponsored by the IEEE Information Theory Group. Ed Posner of Caltech served as the founding general chairman [2].
In its early years, NIPS was a relatively small gathering focused on the intersection of neuroscience and computation. The conference grew steadily through the 1990s and 2000s as neural network research experienced its ups and downs. The deep learning revolution that began around 2012 transformed NeurIPS into a massive event, with attendance growing from approximately 2,000 in 2012 to over 16,000 by 2024.
The conference operated under the acronym NIPS for over three decades. In 2018, following reports of inappropriate behavior at the conference and a letter signed by over 120 academics from Johns Hopkins University and other institutions, the conference board voted to change the name to NeurIPS. The original acronym had become the subject of unwelcome wordplay, and the board determined that a name change would create a more inclusive environment. The new name, NeurIPS, emerged organically from community discussions as the preferred alternative [3].
NeurIPS has served as the venue for many of the most influential papers in the history of machine learning and artificial intelligence. The following table highlights landmark papers presented at the conference:
| Year | Paper | Authors | Significance |
|---|---|---|---|
| 1987 | "Learning representations by back-propagating errors" (related work presented) | Rumelhart, Hinton, Williams | Early NeurIPS helped popularize backpropagation, which became the foundation of neural network training |
| 2001 | "Latent Dirichlet Allocation" | Blei, Ng, Jordan | Foundational topic modeling method that shaped probabilistic machine learning for a decade |
| 2012 | "ImageNet Classification with Deep Convolutional Neural Networks" (AlexNet) | Krizhevsky, Sutskever, Hinton | Launched the deep learning revolution by winning ImageNet with a CNN, reducing error by a dramatic margin |
| 2014 | "Generative Adversarial Nets" | Goodfellow, Pouget-Abadie, Mirza, Xu, Warde-Farley, Ozair, Courville, Bengio | Introduced GANs, one of the most influential generative model frameworks |
| 2014 | "Sequence to Sequence Learning with Neural Networks" | Sutskever, Vinyals, Le | Established the encoder-decoder paradigm for neural machine translation |
| 2015 | "Deep Reinforcement Learning with Double Q-Learning" | van Hasselt, Guez, Silver | Advanced deep RL by addressing overestimation bias in Q-learning |
| 2017 | "Attention Is All You Need" | Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, Polosukhin | Introduced the Transformer architecture, the foundation of virtually all modern LLMs; cited over 173,000 times as of 2025 [4] |
| 2020 | "Language Models are Few-Shot Learners" (GPT-3) | Brown, Mann, Ryder, Subbiah, et al. | Demonstrated that scaling language models enables few-shot learning across many tasks |
| 2022 | "FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness" | Dao, Fu, Ermon, Rudra, Re | Introduced IO-aware attention computation that dramatically reduced memory usage and improved speed for transformer training and inference [5] |
| 2023 | "Direct Preference Optimization: Your Language Model is Secretly a Reward Model" (DPO) | Rafailov, Sharma, Mitchell, Manning, Ermon, Finn | Simplified RLHF alignment by eliminating the need for a separate reward model [6] |
| 2024 | "Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction" | Various | Best Paper at NeurIPS 2024; advanced autoregressive image generation |
NeurIPS follows a multi-track format spanning approximately one week, typically in early to mid-December. The conference structure includes several components:
The core of NeurIPS consists of:
The days following the main conference feature dozens of parallel workshops on specialized topics. Workshops are smaller, more interactive venues that often focus on emerging areas. They have their own submission and review processes, and many important research directions first coalesce at NeurIPS workshops before becoming mainstream.
The day before the main conference features tutorials: half-day or full-day instructional sessions led by experts on specific topics. These are designed to bring attendees up to speed on active research areas.
NeurIPS hosts a series of competitions where participants tackle specific problems, often with standardized benchmarks and evaluation criteria. Notable competition tracks have included:
An industry expo runs alongside the conference, where companies present their AI research, products, and job opportunities. Major technology companies, AI startups, and research labs maintain booths and host presentations. The expo has grown substantially as industry investment in AI has increased.
NeurIPS has experienced dramatic growth in submissions over the past decade, reflecting the explosion of interest in machine learning research:
| Year | Submissions | Accepted | Acceptance Rate | Attendees |
|---|---|---|---|---|
| 2015 | 1,838 | 403 | 21.9% | ~3,800 |
| 2017 | 3,240 | 678 | 20.9% | ~8,000 |
| 2019 | 6,743 | 1,428 | 21.2% | ~13,000 |
| 2021 | 9,122 | 2,344 | 25.7% | ~21,000 (virtual) |
| 2022 | 10,411 | 2,672 | 25.7% | ~10,000 |
| 2023 | 12,343 | 3,218 | 26.1% | ~16,000 |
| 2024 | 15,671 | ~4,000 | 25.8% | ~16,000 |
| 2025 | 21,575 | 5,290 | 24.5% | ~18,000 |
The acceptance rate has remained remarkably stable at approximately 24% to 26% despite the explosive growth in submissions [7]. This stability suggests that the program committee has expanded capacity proportionally rather than becoming more selective. NeurIPS 2025 saw a 61% increase in submissions over 2024, the largest year-over-year jump in the conference's history.
NeurIPS has been held at various locations, predominantly in North America:
| Year | Location | Notes |
|---|---|---|
| 1987 | Denver, Colorado, USA | Inaugural conference |
| 2015 | Montreal, Canada | |
| 2016 | Barcelona, Spain | One of the few non-North American venues |
| 2017 | Long Beach, California, USA | |
| 2018 | Montreal, Canada | First year under the NeurIPS name |
| 2019 | Vancouver, British Columbia, Canada | |
| 2020 | Virtual | Due to COVID-19 pandemic |
| 2021 | Virtual | Due to COVID-19 pandemic |
| 2022 | New Orleans, Louisiana, USA | Return to in-person |
| 2023 | New Orleans, Louisiana, USA | |
| 2024 | Vancouver, British Columbia, Canada | |
| 2025 | San Diego, California, USA (primary) + Mexico City (secondary site) | First year with a dual-location format [8] |
The 2025 introduction of a secondary conference site in Mexico City was a notable experiment aimed at improving accessibility and reducing the environmental impact of travel for Latin American researchers.
NeurIPS uses a double-blind peer review process. Papers are reviewed by members of the program committee, which in recent years has included over 10,000 reviewers. Each paper typically receives 3 to 4 reviews, followed by author rebuttals and a discussion period among reviewers.
The conference has been at the forefront of experimenting with review process improvements:
NeurIPS 2025, held December 2 through 7 in San Diego with a simultaneous site in Mexico City, was the largest edition of the conference to date [8].
Scale: The conference received 21,575 submissions and accepted 5,290 papers (24.5% acceptance rate), with 61 oral papers and 327 spotlight papers.
Best Paper Awards: Seven papers received best paper and runner-up awards, spanning areas including diffusion model theory, self-supervised reinforcement learning, attention mechanisms for large language models, reasoning capabilities in LLMs, online learning theory, neural scaling laws, and benchmarking methodologies for language model diversity.
Key themes: Major trends reflected in the 2025 program included continued emphasis on large language and foundation models, growing interest in reproducibility and data-centric research through the expanding Datasets and Benchmarks Track, and increased attention to the societal impacts of AI.
NeurIPS has played an outsized role in shaping the trajectory of AI research. Several factors contribute to its influence:
Launching paradigm shifts. Many of the most consequential ideas in modern AI were first presented at NeurIPS, including deep convolutional networks for image recognition (AlexNet, 2012), generative adversarial networks (2014), the Transformer architecture (2017), and scaling laws for language models. The conference serves as a launchpad where new ideas receive immediate scrutiny and attention from the global research community.
Industry-academia bridge. Unlike many academic conferences, NeurIPS attracts heavy participation from industry research labs. Google DeepMind, Meta AI (FAIR), Microsoft Research, OpenAI, and other organizations are major contributors. This creates a uniquely productive feedback loop between fundamental research and applied systems.
Benchmark setting. Many standard benchmarks and evaluation protocols in machine learning were introduced or popularized through NeurIPS papers and competitions. The Datasets and Benchmarks track, introduced in 2021, formalized the conference's commitment to rigorous evaluation methodology.
Community formation. NeurIPS workshops have been instrumental in forming new research subcommunities. Areas like AI safety, machine learning for climate, and AI for science were incubated in NeurIPS workshops before becoming established fields with their own dedicated venues.
NeurIPS exists within an ecosystem of top machine learning venues:
| Conference | Focus | Timing | Relationship to NeurIPS |
|---|---|---|---|
| ICML | Machine learning (broad) | July | Co-equal top venue; more focused on core ML methods |
| ICLR | Representation learning | May | Younger conference; strong in deep learning; open review process |
| AAAI | Artificial intelligence (broad) | February | Broader AI scope; less focused on ML specifically |
| CVPR | Computer vision | June | Dominant venue for vision research |
| ACL/EMNLP | Natural language processing | Variable | Dominant venues for NLP-specific work |
As of early 2026, NeurIPS continues to be the primary venue where many breakthrough results in AI are first reported. The conference faces ongoing challenges related to its enormous scale: managing review quality with over 20,000 submissions, ensuring fair and consistent evaluation, and maintaining community culture as the field grows.
The 2025 experiment with dual conference sites may signal a future trend toward more distributed conference formats, addressing both accessibility concerns and the logistical challenges of hosting over 15,000 attendees at a single venue.
With the continued explosive growth of AI research investment and the expanding number of researchers entering the field, NeurIPS is likely to remain at the center of the machine learning research community for the foreseeable future [10].