My NeurIPS 2025 experience

I attended NeurIPS 2025 in San Diego. My focus was mostly on representation learning, physics-informed ML and applications in earth sciences. Here are some of the interesting papers and posters I came across.

Benchmarking

I attended the tutorial on The Science of Benchmarking which was a great overview of best practices.

Several benchmarking papers that were interesting:

• OceanBench: A Benchmark for Data-Driven Global Ocean Forecasting Systems
• Open-Insect: Benchmarking Open-Set Recognition of Novel Species in Biodiversity Monitoring (David Rolnick’s group)
• TreeFinder: A US-Scale Benchmark Dataset for Individual Tree Mortality Monitoring Using High-Resolution Aerial Imagery
• CarbonGlobe: A Global-Scale, Multi-Decade Dataset and Benchmark for Carbon Forecasting in Forest Ecosystems
• MONITRS: Multimodal Observations of Natural Incidents Through Remote Sensing
• AtmosSci-Bench: Evaluating the Recent Advance of Large Language Model for Atmospheric Science
• Mars-Bench: A Benchmark for Evaluating Foundation Models for Mars Science Tasks
• REOBench: Benchmarking Robustness of Earth Observation Foundation Models

Physics-Informed ML

One poster that stood out was Emulator Superiority: When Machine Learning for PDEs Surpasses its Training Data by Felix Koehler & Nils Thuerey (TU Munich). They show that neural networks can outperform the numerical simulator that produced their training data. I think this is a very interesting result for many working on ML for science acceleration.

Earth Observation & Climate

• PhySense: Sensor Placement Optimization for Accurate Physics Sensing
• Partial Physics Informed Diffusion Model for Ocean Chlorophyll Concentration Reconstruction
• FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution
• Global 3D Reconstruction of Clouds & Tropical Cyclones - Spotlight at the Tackling Climate Change with ML workshop
• Mass Conservation on Rails – Rethinking Physics-Informed Learning of Ice Flow Vector Fields - presented at the Tackling Climate Change with ML workshop
• SHRUG-FM: Systematic Handling of Real-world Uncertainty for Geospatial Foundation Models

Vision & Representation Learning

REOrdering Patches Improves Vision Models (UC Berkeley & U Pittsburgh) - they showed that patch ordering matters for vision transformers and developed an optimal reordering algorithm.

Also enjoyed seeing the BikeBench poster, a bicycle design benchmark for generative models with objectives and constraints. Fun to see ML applied to physical design problems.