LLMs can ace science exams and explain complex mechanisms. But can they actually do science—design a molecule, a dosing regimen, or a gene circuit that works? We built SciDesignBench, a benchmark for testing frontier LLMs on scientific design tasks grounded in real simulators. The core finding is that models that can talk fluently about science still struggle when asked to produce designs that actually satisfy quantitative targets. We also propose a simulator-feedback RL recipe that improves performance on these tasks. Paper: arxiv.org/abs/2603.12724 🧵

The elegant part is that the same asymmetry that makes inverse design hard also suggests the fix. Scientists have already built forward simulators for many of these problems. Those simulators can be repurposed as RL environments: propose a design, evaluate it, learn from the reward. In our case studies, training on simulator feedback lets a small 8B model outperform much larger frontier models on selected scientific design tasks.

We teach models to reason about biology. Today we're open-sourcing the tool we built to run that science. Think Claude Code, but for drug discovery. pip install celltype-cli

We benchmarked celltype-cli against other agentic systems on computational biology tasks.