Analysis of NeurIPS 2025 Papers

The best papers from the roughly 5,000 submissions to NeurIPS that received Best Paper Awards have now been published. For the final session of our seminar, we decided to give a brief review of them.

There will be seven (or perhaps six—we’ll see) talks, each lasting 10–15 minutes:

— Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)

— Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free

— 1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities

— Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training

— Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

— Optimal Mistake Bounds for Transductive Online Learning

— Superposition Yields Robust Neural Scaling