Ben Blaiszik

You can grow fresh peaches in your yard as far north as Chicago. Quite the unlock.

Thanks for your comments Evan, will read more on these topics.

Hi Evan, in the other year's events, we have had projects spanning from data organization to improving clean water access, increasing access to energy storage and collection, and limiting material toxicity. All of these and others are shared openly in our papers. Do you have suggestions?

Please follow for updates. Excited to meet everyone this year!

What's coming in the next few weeks? Site announcements, speaker announcements

Register or join the Slack here: llmhackathon.github.io

🌐 Meet and collaborate with amazing people

📝 Gain experience building applications, to further your research and build your resume

🙌 Participate fully online or attend an in-person site

🖊️ Earn authorship on an event summary paper. Prior submissions have already been cited hundreds of times!

We're about a month and a half away from the 3rd LLM Hackathon for Applications in Materials and Chemistry (Sept 11-12). I hope you will join us! Here are a few reasons to register now.

🤗 Join the preeminent community for LLM and multimodal model applications in materials science and chemistry.

Learning from one and only one shot npj Artificial Intelligence - Learning from one and only one shot

There remains significant work to port this to more complex images, but this is a great step forward, and provides a path for others to continue.

James Evans, Lav Varshney, Igor Mineyev, Haizi Yu.

Paper: rdcu.be/ev44p

A key insight. Humans naturally "distort" objects mentally when comparing them - and this model is built to do the same.

Using this approach, they achieve near-human performance on character/doodle recognition without transfer learning and with one or few examples.

a man is singing into a microphone in front of a crowd at a concert . ALT: a man is singing into a microphone in front of a crowd at a concert .

Training data volumes have exploded, but what if we only get one shot to teach a model?

Evans et al show a breakthrough: a "distortable canvas" model that mimics human cognition, achieving 80% MNIST accuracy with just 1 example per class. No pretraining needed.

The first (of many, I hope!) challenges articles is finally out! Really excited to be able to provide this service to the #machinelearning community. If you're organizing #hackathons, #MLcompetitions, or similar, please reach out! @iopp-mlresearch.bsky.social #MachineLearningSeries

LLM Hackathon for Applications in Materials and Chemistry · Luma Join us for an extraordinary convergence of minds where artificial intelligence meets scientific discovery. Building on our previous hackathons, we're…

Ready to go? Sign up now!

Event Registration (free): lu.ma/hspoki8y

Slack for teaming and ideation: join.slack.com/t/llmsformat...

Paper w/ last year’s outputs: arxiv.org/abs/2505.03049

How can you participate?
- In-person sites around the world (>10 sites)
- A full-featured virtual option enabling people to participate from around the world. If you have an internet connection, you can participate.
- Join our Slack for access to a community of over 500 LLM practitioners

Solving the world's biggest problems requires advancing our capabilities in materials & chemistry faster than ever. This is your opportunity to help build the applications that will transform the future and to join a community that will uplift you and provide a path to gain valuable AI experience.

🚀 Announcing the third LLM Hackathon for Applications in Materials and Chemistry (Sept 11-12)!

Last year, 34 teams submitted inspiring and open examples. This year, we are expecting amazing applications with higher powered models and agentic frameworks. Your imagination is the only limit.

➡️ You can implement this approach with a single, reusable system prompt to deliver a deeper “chain-of-reasoning”.

Wonderful work by the team at Argonne, UIC, and NU including Hassan Harb, Rajeev Assary, Brian Ingram and others.

Paper link: doi.org/10.26434/che...
Prompt: github.com/HassanHarb92...

Why does this matter?

➡️ The prompted LLM saw improved performance, achieving SOTA (97.15%) on the ARC Challenge without fine-tuning or external tool usage.

➡️ The approach showed consistent gains in reasoning depth, clarity, and domain-specific insight across chemistry and materials science.

Figure 2: Accuracy comparison on the ARC Challenge dataset. The SM Agent (this work, shown in pink) achieves 97.15% overall accuracy, outperforming all prior public submissions (shown in orange), including GPT-4o+CoT (94.88%). Unlike other top entries, which often rely on ensemble methods or retrieval-augmented inputs, the SM Agent uses only structured prompting via a single Socratic system prompt. Results of previous models are obtained from the AI2 ARC leaderboard.

A new preprint introduces a Socratic Method Agent that guides LLMs through structured questioning, i.e., definition, analogy, hypothesis elimination, and more following lessons learned from thousands of years of philosophy and reasoning achieving SOTA on the ARC Challenge.

Continued below w links:

Son proudly and triumphantly: “We really can accomplish anything”

Me: 🥺🥲🫡

Our accomplishment: Getting a Nintendo Switch 2

Merged community requests from Oliver Schilter and @mlederbauer.bsky.social. Thank you. 🙌

Other minor: There is now a changelog with relevant info about each dataset. It is generated by AI, but seems good for a changelog.

Even more open materials and chemistry data for everyone! Think of all the models you could train using these datasets.

June updates: 26 high-quality datasets spanning polymer science, drug discovery, spectroscopy, MOF databases, + foundation model training datasets.

🔗: github.com/blaiszik/awe...

The talk touches on many topics, including:

🔷 What is an agent?
🔶 Why do we want them?
🔷 How can agents accelerate research?
🔶 How can agents work in tandem with humans?

YouTube video by alphaXiv AI4Science w: Sam Schmidgall

Are you interested in learning about how AI agents could be used in various research tasks? In this talk, Sam Schmidgall discusses "Autonomous Research Assistants". The talk is very accessible for those looking to get started.

📽️ Video: www.youtube.com/watch?v=5lp-...

a cartoon of patrick star meditating with the words manifesting mode engaged written below him ALT: a cartoon of patrick star meditating with the words manifesting mode engaged written below him

Manifesting

Nintendo Switch affixed with a piece of paper with the number two written on it.

Someone is trying to drop me a subtle hint… #Switch2

Ganz gut. 😎✨ Viel Wissenschaft kommt noch

Paired with the new Aurora system at ALCF, I can't wait for the new discoveries, brought to you by the Department of Energy.

Article with more detail: www.anl.gov/article/adva...

America and the world's brightest X-ray "microscope" - now 500x brighter. It can:

🔋 watch a battery charge atom by atom,
🛫 help us understand fracture in fighter jet alloys, and
🧬 speed up cancer-drug discovery.

Welcome to the new Dept. of Energy Advanced Photon Source at Argonne! 🥳🔬⚡️

Looking for a place to publish your datasets, Check out @DataFacility.

next.materialsdatafacility.org

GitHub - blaiszik/awesome-matchem-datasets Contribute to blaiszik/awesome-matchem-datasets development by creating an account on GitHub.

Know of other great datasets that should be added to the Awesome Materials and Chemistry Dataset List? Issue a PR to the repo, or reply here.

Link: github.com/blaiszik/aw...

GitHub - blaiszik/awesome-matchem-datasets Contribute to blaiszik/awesome-matchem-datasets development by creating an account on GitHub.

We've also added many more experimental and computational datasets! including the recent OMol25 (100M+ structures).

Awesome Materials and Chemistry Dataset Link: github.com/blaiszik/aw...

x.com/mshuaibii/s...