A recent arXiv paper introduces Anchor-and-Harvest Curriculum (AHC), a training system designed for the Craftax-Full environment. Unlike approaches that lean on large language models, AHC is described as a fully unmasked, non-LLM curriculum built from domain-specific rules.

The method pairs achievement-oriented tasks with harvested initial-state progress, allowing the curriculum to adapt difficulty based on learner performance. It also incorporates bounded selection to keep training focused. The abstract highlights these components as the core of the proposed system, though details on implementation and results are not included in the available text.

This work is notable for offering a rule-based alternative to LLM-driven curriculum generation, potentially making reinforcement learning in complex environments more accessible and interpretable.