Non-LLM curriculum outperforms LLM-generated methods in Craftax-Full

Researchers have developed the Anchor-and-Harvest Curriculum (AHC), a novel training system for the Craftax-Full environment that does not rely on large language models (LLMs). AHC utilizes achievement-oriented tasks, adaptive difficulty, and inherited progress to train agents. In evaluations using 2,005,401,600 PPO rollout transitions, AHC achieved a native-task return of 46.70 ± 0.57, outperforming a standard PPO-GTrXL model which scored 41.91 ± 1.01. The results suggest that explicit, non-LLM curricula can achieve performance comparable to LLM-generated curricula in complex environments. AI

IMPACT Demonstrates that explicit, non-LLM curricula can achieve competitive performance in complex environments, potentially reducing reliance on large language models for training.

RANK_REASON The cluster describes a new research paper detailing a novel training curriculum for an AI agent. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Anchor-and-Harvest Curriculum
• Craftax-Full
• DICODerma: A Practical Approach for Metadata Management of Images in Dermatology
• GTrXL
• Proximal Policy Optimization

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →