Early this morning, NeurIPS 2025, the world's leading artificial intelligence conference, announced its key annual awards. These include the Best Paper Award, Runner-up Best Paper, Test of Time Award, and the Sejnowski-Hinton Award.

This year, the conference selected 4 best papers and 3 runner-up papers, with research areas spanning key frontiers such as diffusion model theory, attention mechanisms, new reinforcement learning paradigms, online learning theory, and large model diversity characterization. Additionally, the classic paper , which has had a profound impact on computer vision over the past decade, received the Test of Time Award, marking a historic milestone at the conference.

NeurIPS 2025 received 21,575 submissions, with 5,290 papers accepted, yielding an acceptance rate of just 24.52%.

Below is a detailed analysis of this year's most notable achievements.

The first paper, titled , is from Stanford University.

The research team introduced Infinity-Chat, a new large-scale dataset comprising 26,000 open-ended real-world queries, and constructed, for the first time, a classification system for open-ended LLM tasks. Through systematic evaluation across different models, the authors uncovered two key phenomena:

• Intra-model repetition: The same model repeatedly generates similar responses.
• Inter-model homogeneity: Different models produce highly similar outputs.

The authors vividly described this as the "Artificial Hivemind"—a pervasive pattern collapse effect in open-ended generation tasks.

Beyond proposing the dataset and evaluation framework, the paper directly addresses core issues of modern LLMs, including "diversity," "value pluralism," and "social impact." Reviewers praised it as offering "a new foundational understanding of real-world AI behavior."

Paper link: https://openreview.net/pdf?id=saDOrrnNTz

The second paper, titled , is a collaborative effort between scholars from Alibaba's Tongyi Qianwen and the University of Edinburgh.

This paper from Alibaba's team is the first to systematically demonstrate the critical role of gating mechanisms in attention and directly applied these findings to the latest Qwen3-Next model.

Key contributions include:

• Proving the sources of gating-enhanced performance: Enhanced nonlinearity and input-dependent sparsity.
• Resolving model issues such as Attention Sink and Massive Activation.
• Significantly improving training stability and reducing Loss Spikes.
• Dramatically enhancing long-context extrapolation capabilities.

Paper link: https://openreview.net/forum?id=1b7whO4SfY

The third paper, titled <1000 Layer Networks for Self-Supervised RL>, comes from Princeton University and the Warsaw University of Technology.

Reinforcement learning (RL) has long been considered ineffective for training extremely deep networks. However, this paper directly scales network depth to 1,024 layers, achieving performance improvements of 2×–50× across multiple tasks.

The paper demonstrates that increasing depth not only boosts success rates but even fundamentally alters agent exploration behavior. This work challenges the long-held implicit assumption that "RL is unsuitable for deep networks" and is poised to become a key paradigm for future RL scaling.

Paper link: https://openreview.net/pdf?id=s0JVsx3bx1

The fourth paper, titled , is from PSL Research University and Bocconi University.

Why do diffusion models maintain strong generalization despite large-scale training? This paper provides a robust theoretical explanation.

First, the model exhibits two critical time points:

• t_g: The onset of generating high-quality samples.
• t_m: The onset of overfitting and memorization.

t_m delays linearly with increasing data scale, while t_g remains largely unchanged. Thus, a safety margin expands as data grows. This mechanism constitutes an implicit dynamic regularization.

Paper link: https://openreview.net/pdf?id=BSZqpqgqM0

The first runner-up paper, titled , comes from Tsinghua University and Shanghai Jiao Tong University.

This paper overturns industry consensus by demonstrating that RLVR (RL with verifiable rewards) does not truly enable models to acquire new reasoning capabilities. The findings challenge a major mainstream approach to enhancing LLM reasoning over the past two years.

Paper link: https://openreview.net/pdf?id=4OsgYD7em5

The second runner-up paper, titled , comes from Kent State University and Purdue University.

This paper resolves a 30-year theoretical challenge by providing the optimal mistake bounds for transductive online learning: a lower bound of Ω(√d) and an upper bound of O(√d).

The two bounds perfectly match, forming a tight result. This achievement is regarded as a textbook-level theoretical breakthrough.

Paper link: https://openreview.net/pdf?id=EoebmBe9fG

The third runner-up paper, titled , comes from MIT.

The research team proposes that representation superposition is the fundamental source of neural network scaling laws.

This work explains at a mechanistic level why model loss decreases as a power law when models scale up, serving as a crucial piece in understanding scaling laws.

Paper link: https://openreview.net/pdf?id=knPz7gtjPW

The Test of Time Award was presented to the paper , authored by Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun.

A decade ago, this paper achieved a unification of extremely high accuracy and near-real-time (5 FPS) detection, enabling neural network object detection models to be truly deployed in real-world applications.

To date, the paper has been cited over 56,700 times, and its framework continues to evolve and find applications in numerous tasks.

Paper link: https://arxiv.org/pdf/1506.01497

The Sejnowski-Hinton Award was presented to the 2016 paper , authored by Timothy Lillicrap, Daniel Cownden, Douglas Tweed, and Colin Akerman.

This paper introduced the highly influential feedback alignment mechanism, demonstrating that neural networks can approximate backpropagation without precisely symmetric feedback weights, enabling effective training through local learning rules.

This mechanism has spurred extensive research on biologically plausible learning.

Notably, the award itself embodies a legendary academic tale—the 40-year friendship between Hinton and Sejnowski, along with their promise to share Nobel Prize winnings, ultimately led to the establishment of this award.

Paper link: https://www.nature.com/articles/ncomms13276

References:

https://mp.weixin.qq.com/s/4Jb_jiLQwrs7GK9V7NOjAg

https://blog.neurips.cc/author/mengyeren/

https://blog.neurips.cc/2025/11/26/announcing-the-2025-sejnowski-hinton-prize/