NeutronGym Benchmark Tests LLM Agents on Physics-Based Instrument Design

NeutronGym provides an executable environment for scientific instrument design using McStas ray-tracing for non-LLM grading. Reinforcement learning on this reward signal improved Qwen3-8B performance from 11% to 77% on held-out design instances.