Its always nice to meet old (and a lot of new!) friends!
Last week I talked at @dcl-ethz.bsky.social for their Data+ML lunch about our latest research (arxiv.org/abs/2503.08537)
Thank you so much @spschmid.bsky.social and @valencekjell.com for the invitation, first time for me at @ethz.ch 🤩

Its always nice to meet old (and a lot of new!) friends!
Last week I talked at @dcl-ethz.bsky.social for their Data+ML lunch about our latest research (arxiv.org/abs/2503.08537)
Thank you so much @spschmid.bsky.social and @valencekjell.com for the invitation, first time for me at @ethz.ch 🤩

Here's a small gif illustrating the idea 😋

We made a small benchmark for this task **steerable synthesis planning**, with several targets and prompts, and the results are quite impresive!

Larger models are quite good at the task 🚀, with newer versions improving rapidly. A small model like gpt-4o-mini is not as good tho

4/8

In synthesis planning, we input mol. and the output are some routes, with no restrictions. But what if I want to specify details? - see img
In our work, users can specify a query, and during search the LLM selects only aligned solutions - working as a zero-shot value function!
3/8

LLMs are pretty bad at writing molecules, but quite good at analyzing mols and reactions!
In our new work we use LLMs+search in chemical tasks, unlocking steerable synth. planning and mechanism prediction 🌟

Chehck out the paper 👉 arxiv.org/abs/2503.08537

1/8

Back in May, we hosted an on-site hub here in #Lausanne at @epfl-ai-center.bsky.social, generously supported by #LIAC (the @pschwllr.bsky.social group), and our sponsors Atinary, EPFL AI Center and El Tony Mate. Thank you very much Loïc Roch and Hermann Tribukait for your invaluable support 🫶!