Reproduce “XLand-100B: A Large-Scale Multi-Task Dataset for In-Context Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alexander Nikulin, Ilya Zisman, Alexey Zemtsov, Vladislav Kurenkov
XLAND-100B is a new large-scale, multi-task dataset for in-context reinforcement learning (ICRL), providing over 100 billion transitions across nearly 30,000 tasks, each containing complete learning histories with improving policies. This dataset addresses the critical scarcity of ICRL-compatible data, aiming to accelerate the development of generalist agents and enable the study of scaling laws.
Key insights
- Large-scale datasets with complete learning histories are essential for the emergence and scaling of in-context learning abilities in reinforcement learning agents.
- Current in-context RL methods, such as Algorithm Distillation, demonstrate initial ICL capabilities but struggle significantly with increased task complexity.
- Partial observability in environments poses a fundamental challenge for certain ICRL architectures like Decision-Pretrained Transformers, which require explicit state information in context.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.