LLMs are autoregressive and slow? No! Parallel Token Prediction decodes multiple consistent tokens in one model call. PTP allows arbitrary dependencies in one call, unlike discrete diffusion. Practical: 2.4x speedup
github.com/mandt-lab/ptp
ICLR: Apr 23, morning poster P3-#608
If you’re curious about my research, feel free to check out the paper. I’m always happy to chat, so please don't hesitate to reach out!
📑 Paper: www.pnas.org/doi/10.1073/...
Huge thanks to the Center for Climate Systems Modeling (C2SM) and the EXCLAIM team for organizing such a well-curated event with an amazing agenda. It was an honor to be part of it.
🧵 7/n
📊 Harnessing observations is critical. Since even high-resolution models are imperfect, effectively leveraging observational data through AI-empowered data assimilation becomes crucial for improving model accuracy and reliability.
🧵 6/n
⚙️ Differentiable physics is essential. There’s a pressing need to redesign physics-based models so that they are differentiable, enabling seamless integration with AI frameworks.
🧵 5/n
Here are some of my key takeaways:
🔗 Hybrid approaches are the future. Rather than relying solely on physics-based models or purely data-driven approaches, the real breakthrough lies in combining both approaches to unlock the best of both worlds.
🧵 4/n
The symposium featured an inspiring lineup of talks that showcased the massive developments happening at the intersection of AI and climate science.
🧵 3/n
The poster session was exciting. I had the chance to connect with fantastic researchers, many of whom showed genuine interest in my work and engaged in meaningful discussions.
🧵 2/n
Just got back from an incredible experience at the EXCLAIM symposium “Is AI the Future of Weather and Climate Modeling?” at the @ethz.ch.
🧵 1/n
What an exciting collaboration!
Congratulations, @f-immorlano.bsky.social @gabrieleaccarino.bsky.social @stephanmandt.bsky.social @pierregentine.bsky.social + colleagues at CMCC Foundation, @columbiauniversity.bsky.social @ucirvine.bsky.social @dlr-spaceagency.bsky.social @dlr-en.bsky.social
@pierregentine.bsky.social @stephanmandt.bsky.social @leapstc.bsky.social @columbiauniversity.bsky.social @ucirvine.bsky.social
Truly grateful to my mentors - Pierre Gentine, Giovanni Aloisio, and Stephan Mandt - and to the co-authors for their support and the opportunity to work on such an amazing project at the intersection of AI and climate. 🧵 4/4
In this work, we use a Transfer Learning strategy to constrain CMIP6 simulations to historical observations, ultimately providing more precise and reliable estimates of global warming spatially resolved on global and regional scales. 🧵 3/4
Earth System Models are currently the main tools used to project climate change according to future emission pathways. However, they still exhibit large uncertainties which are a major roadblock for policymakers. 🧵 2/4
I am so excited to share our latest research paper "Transferring climate change physical knowledge" published in @pnas.org! 🧵 1/4
www.pnas.org/doi/10.1073/...
So delighted to share our new paper, my little Covid project brought to a whole new level by Francesco Immorlano www.pnas.org/doi/10.1073/.... We demonstrate how future projections can be refined using transfer learning (fine tuning). This refines our estimates of global warming.
Our Dep. Dir. Tian Zheng connected w/Cali LEAPers @jlin404.com + @f-immorlano.bsky.social and other colleagues (Katherine Frields, Liran Peng, Yan Xia) while at an event near @ucirvine.bsky.social ... Just proves: science = more fun together! 💪
#LEAPcolors #community #ML #AI #physics #data #climate
Excited to share some great collaboration on a new data assimilation framework with transformer encoding arxiv.org/pdf/2502.02884. Exciting times for data assimilation with ML
New postdoc position open at UCI to work at the intersection of machine learning and climate science, jointly between Stephan Mandt's group and my group.
More details and application information here: ics.uci.edu/~smyth/postd... Feel free to share.
"Extreme wet-to-dry transitions can amplify wildfire risk by allowing increased rates of plant growth to be immediately followed by rapid drying of flammable vegetation, increasing the potential intensity of subsequent fire events"
www.nature.com/articles/s43...
@weatherwest.bsky.social
group photo at UCI Social event, NeurIPS 2024
Faculty + guest at UCI Social at NeurIPS 2024
Pictures from the recent UCI Social at @neuripsconf.bsky.social with a subset of students, faculty, guests. Looking forward to the next event at NeurIPS 2025 in San Diego!
We spy @f-immorlano.bsky.social + LEAP alum Sungduk Yu, who supported @juannat7.bsky.social @ #NeurIPS2024: "ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction."
#NSF #CUSEAS
@neuripsconf.bsky.social
@columbiaclimate.bsky.social
@ucirvine.bsky.social
Hello (climate-ML) world!
Can you please add me? Thank you so much!
@f-immorlano.bsky.social
