MC-EvoBench: Auditing Experience-Driven Self-Improvement in Minecraft Agents
MC-EvoBench tests whether Minecraft agents convert experience into durable capabilities that transfer across reset worlds.
Researchers introduce MC-EvoBench, a controlled Minecraft benchmark with 19 registered evolution chains, including 12 primary evaluation chains and seven calibration or probe chains. Tasks are represented as task-state graphs with stage-level verifiers, and only a declared agent artifact may cross a reset source-target boundary. Related experience showed no stable transfer, with confidence intervals including zero, while early wood-collection and crafting bottlenecks often left later capabilities unobserved. Code and data are released with the paper.
- Nineteen evolution chains, with 12 primary and seven probe chains
- Only a declared agent artifact crosses a reset world boundary
- Related experience did not stably improve either tool probe
- Early crafting bottlenecks leave later skills unobserved
From the original article
1 min read at arxiv.orgA central challenge in embodied recursive self-improvement is enabling agents to convert interaction history into capabilities that persist beyond one world and transfer to later tasks.
By Jingwei Song · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.