The annual Conference on Neural Information Processing Systems (NeurIPS) has announced its 2025 Best Paper Awards, recognizing groundbreaking research that pushes the boundaries of artificial intelligence. Below are summaries of the award-winning and runner-up papers, highlighting their key contributions and a new study on LLM homogenization.
The “Artificial Hivemind”: LLMs Are Converging on a Single Mind
Authors: Liwei Jiang, Yuanjun Chai, Margaret Li, Mickel Liu, Raymond Fok, Nouha Dziri, Yulia Tsvetkov, Maarten Sap, Yejin Choi
A new study reveals a phenomenon termed the “Artificial Hivemind,” where modern large language models (LLMs) are increasingly producing identical or highly similar outputs. The researchers introduced INFINITY-CHAT, a dataset of 26,000 real-world, open-ended questions, to systematically evaluate the output diversity of over 70 state-of-the-art LLMs.
The findings demonstrate an extreme “mode collapse” in which models not only repeat the same answers internally but also converge on strikingly similar responses across different model families. This homogenization challenges the common assumption that techniques like increasing temperature or using model ensembles can guarantee diverse outputs. The research suggests that modern training methods, such as Reinforcement Learning from Human Feedback (RLHF) and instruction tuning, have so narrowed the creative potential of LLMs that distinct models like DeepSeek and GPT-4 often behave as near-identical clones. Furthermore, the study indicates that current reward models are poorly equipped to value pluralism, failing to correctly score valid but unconventional human-preferred responses.
Gated Attention: A Simple Fix for Stable and Efficient Transformers
Authors: Zihan Qiu, Zekun Wang, Bo Zheng, Zeyu Huang, Kaiyue Wen, Songlin Yang, Rui Men, Le Yu, Fei Huang, Suozhi Huang, Dayiheng Liu, Jingren Zhou, Junyang Lin (Qwen Team)
Researchers have introduced Gated Attention, a simple architectural modification that significantly improves the stability and performance of large-scale language models. The mechanism applies a learnable, input-dependent gate to the output of the standard attention mechanism, introducing element-wise sparsity and non-linearity.
This seemingly minor change yields profound benefits. It eliminates the notorious “loss spikes” that plague large-scale training, leading to greater stability and consistently improved perplexity in both Mixture-of-Experts (MoE) and dense models. Crucially, Gated Attention mechanistically resolves the “Attention Sink” phenomenon without requiring heuristic fixes like special sink tokens, thereby enhancing the model’s ability to handle long contexts.
Beyond the Limit: Scaling Reinforcement Learning Policies to 1,000 Layers
Authors: Kevin Wang, Ishaan Javali, Michał Bortkiewicz, Tomasz Trzciński, Benjamin Eysenbach
Challenging the long-held belief that Reinforcement Learning (RL) does not benefit from network depth, this work successfully scales RL policies from the standard 2-5 layers to over 1,000. The breakthrough was achieved by combining Self-Supervised Learning, specifically Contrastive RL, with modern architectural components like residual connections, LayerNorm, and Swish activations.
While conventional RL algorithms like Soft Actor-Critic (SAC) see performance stagnate or collapse with deeper networks, this new approach demonstrates continued performance gains of 20x to 50x. The deep networks enabled agents to solve complex, long-horizon humanoid maze tasks and develop emergent locomotor skills, all without the need for explicit reward engineering. This research opens the door for applying the benefits of deep architectures to a new class of complex RL problems.
Best Paper Award: Why Diffusion Models Generalize Instead of Memorize
Authors: Tony Bonnaire, Raphaël Urfin, Giulio Biroli, Marc Mézard
This award-winning paper provides a theoretical and empirical explanation for a central paradox in generative AI: why overparameterized diffusion models produce novel creations rather than simply copying their training data. The authors identify two distinct timescales in the training process: a generation timescale, when the model learns to create valid samples, and a memorization timescale, when it begins to overfit and reproduce specific training examples.
The key insight is that the memorization timescale grows linearly with the size of the dataset, while the generation timescale remains constant. This proves that having a larger dataset naturally creates a wider training window where the model can learn to generalize robustly before it starts to memorize. The work establishes that “early stopping” is not merely a practical heuristic but a fundamental necessity driven by a principle the authors call “Implicit Dynamical Regularization,” explaining how massive models can generalize so effectively.
Best Paper Runner-Up: Reinforcement Learning Refines, But Does Not Expand, LLM Reasoning
Authors: Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Yang Yue, Shiji Song, Gao Huang
This study challenges the idea that reinforcement learning can teach large language models to discover “superhuman” reasoning abilities. Researchers systematically investigated the limits of LLMs trained with Reinforcement Learning with Verifiable Rewards (RLVR), a method that rewards models for producing verifiably correct answers.
Using unbiased metrics across mathematics, coding, and visual reasoning, they found that while RLVR makes models significantly more efficient at finding correct solutions, it does not expand their fundamental reasoning capabilities. In fact, when given enough attempts, the base pre-trained models often solved more unique problems than their RL-tuned counterparts. The findings suggest that current RL methods are bounded by the knowledge and reasoning patterns already present in the pre-trained model, acting as a powerful amplifier rather than a source of novel discovery.
Best Paper Runner-Up: Solving a 30-Year-Old Problem in Online Learning Theory
Authors: Zachary Chase, Steve Hanneke, Shay Moran, Jonathan Shafer
In a significant theoretical breakthrough, researchers have resolved a 30-year-old open problem in learning theory by establishing the precise mistake bounds for Transductive Online Learning. This setting involves predicting a sequence of labels when the unlabeled test points are known in advance.
The paper proves that for a hypothesis class with a Littlestone dimension of d, the optimal number of mistakes an algorithm can make is tightly bound by the square root of d. This result formally quantifies the benefit of “looking ahead,” showing that access to the future sequence of tasks allows for a quadratic reduction in errors compared to the standard online setting, where the bound is linear in d. The work closes a massive exponential gap between previously known upper and lower bounds, providing a definitive answer to a foundational question in the field.
Best Paper Runner-Up: A Geometric Explanation for Neural Scaling Laws
Authors: Yizhou Liu, Ziming Liu, Jeff Gore
This paper offers a new, first-principles explanation for neural scaling laws—the observation that model performance improves predictably with increased size. The authors link this phenomenon to “representation superposition,” a regime where models learn to represent far more features than they have available dimensions by compressing them into a dense space.
By analyzing open-source LLMs, they demonstrate that when models operate in this highly compressed state, performance loss scales inversely with model width. This scaling behavior is not primarily driven by the statistical properties of the data but by the geometric interference between feature vectors packed into the model’s limited dimensions. The crucial implication is that simply scaling up data may not be enough to overcome current performance plateaus; instead, future progress may depend on new architectures designed to more effectively manage this geometric feature interference.
