This workshop is for researchers interested in using computers to find, build, and discover things. It must be possible for a computer to know when it has succeeded and for it to know when it is close.
We encourage attendees from mathematical and physical sciences, and engineering. Graduate students, postdoctoral researchers, and faculty are all welcome. You do not need to be at Yale.
Here are some examples of the kinds of problems these tools can tackle. If your research involves anything like these, this workshop is for you.
If you want to prove an inequality f(x) ≤ 0, you can set up a search for x* such that f(x*) > 0. Finding such a counterexample disproves the conjecture. ShinkaEvolve can search complex, high-dimensional spaces for these rare witnesses far more effectively than random sampling or conventional optimization.
LLM-guided search has been used to find graphs with unusual properties — large cuts, large independent sets — that serve as gadgets in hardness-of-approximation proofs. Brute force search becomes impractical even for small graphs; evolutionary search finds them.
Given an initial model and an evaluation metric — say, validation accuracy on a domain-specific dataset — ShinkaEvolve can iteratively modify the architecture, adding layers, changing connectivity, and trying different training strategies to improve performance.
ShinkaEvolve can evolve algorithms that decompose molecular graphs into structural motifs, improving the accuracy of downstream models that predict properties like toxicity or biological activity.
Given an initial ODE solver, ShinkaEvolve can propose modifications to the integration strategy — alternative operator splittings, time discretizations, interpolation schemes — to reduce numerical error or improve stability.
ShinkaEvolve can search over circuit templates which find the ground state energy of a quantum system or which realize desired operators, optimizing both the circuit structure and its parameters simultaneously.
We are speaking with researchers in the biological sciences about relevant applications. Check back soon.
All software is pre-installed on Yale's Bouchet research computing cluster. Participants receive a temporary account — just SSH in and start working. No local setup required on the day. Or, bring your laptop and hack on things with your personal setup.
Temporary Bouchet accounts provisioned for all registered participants. ShinkaEvolve and Claude Code pre-installed and ready to use.
Each participant receives an individual API key with a spending limit, pre-configured in the cluster environment.
Participants without an existing Claude Pro or Max subscription will receive a one-month gift subscription (~$20) prior to the event. You will be asked about your subscription status at registration.
Two YCRC staff members present for the full day to troubleshoot any technical issues.
Registration closes a few days before April 24 to allow time for account provisioning.
Researcher at Sakana AI and lead developer of ShinkaEvolve. Rob will join the workshop remotely to give a keynote talk. His work focuses on open-ended evolutionary search and making these tools accessible to the research community.
Lead author of AlphaEvolve. Alex’s research at Google DeepMind has shaped how the field thinks about AI-guided scientific discovery.
Opening session: Optimization for Design and Discovery