2 x papers at ICML 2025
We have 2 papers published at ICML 2026:
- Causal Preference Elicitation
- Multi-Scale Wavelet Transformers for Operator Learning of Dynamical Systems
Multi-Scale Wavelet Transformers for Operator Learning of Dynamical Systems
AI models that simulate physics (like weather or climate) tend to blur out fine details, which causes errors to snowball over time. This paper introduces a transformer model that processes data in a “wavelet” format, a way of representing information at multiple zoom levels at once, so it can keep track of both the big picture and the fine details. Tested on chaotic systems and real climate data (ERA5), it produces more accurate, longer-lasting forecasts.
Causal Preference Elicitation
When scientists try to figure out what causes what (e.g., which gene affects which other gene), they often have a rough guess but need an expert to confirm the tricky parts. This paper presents a method that intelligently picks which questions to ask an expert, so their answers narrow down the right cause-and-effect map as fast as possible. It’s tested on gene and protein data, where it recovers the true causal relationships using far fewer expert queries than existing approaches.