MC-EvoBench: Auditing Experience-Driven Self-Improvement in Minecraft Agents
The benchmark finds no stable evidence that related Minecraft experience transfers into reusable agent capability.
MC-EvoBench tests 19 controlled evolution chains and only lets a declared agent artifact carry over between source and target worlds. In the reported probes, related experience does not improve either tool task, and the small wood-species gain depends on one seed with intervals that include zero. The paper also finds that early wood-collection and crafting failures can prevent later capabilities from being observed at all. ArXiv · AI/CL/LG's note
MC-EvoBench tests 19 controlled evolution chains and only lets a declared agent artifact carry over between source and target worlds. In the reported probes, related experience does not improve either tool task, and the small wood-species gain depends on one seed with intervals that include zero. The paper also finds that early wood-collection and crafting failures can prevent later capabilities from being observed at all. ArXiv · AI/CL/LG's note
score 5