Mdforge

“We build MDForge, an LLM agent whose in-context update rule densifies the sparse reward via a multi-agent debate among physics experts.”