Building embodied agents that genuinely learn and transfer skills is harder than it looks—current methods show minimal stable transfer, and early bottlenecks often prevent agents from even attempting later tasks.
MC-EvoBench is a Minecraft benchmark designed to rigorously test whether AI agents can learn reusable skills from experience that transfer to new tasks. It uses controlled task chains, environment resets, and careful measurement to separate genuine capability learning from lucky outcomes or resource carryover, revealing that current agents struggle to build durable, transferable skills.