Pretty clever idea from @allmeasures and @Andrewdblevins at @leashbio to use a dense combinatorial set of experimental data like a DEL screen as an environment to test how well agents can optimize in chemical space. Medchem optimization in the real world isn't so constrained, but it's plausible to me that ability at the DEL task would transfer to the real world. I wonder where else the idea can be applied to build RL environments for science. Blog post: leashbio.substack.com/p/maki… Preprint: github.com/Leash-Labs/minos-… The preprint nicely addresses the biggest question I had from reading just the blog post by performing a "shuffle" ablation suggesting that it was measuring actual chemical reasoning ability instead of just "Mastermind" ability. I would also be interested in seeing some other baselines beyond random, like a fingerprint-based optimizer.