SKILL0
A research framework for in-context reinforcement learning that internalizes reusable skills into an agent's policy, evaluated on tasks such as ALFWorld and Search-QA.
SKILL0 is an in-context reinforcement learning framework for skill internalization: rather than storing skills externally, it trains an agent through reinforcement learning so that useful behaviors become part of its in-context policy. The paper reports substantial improvements over a standard reinforcement learning baseline on the ALFWorld embodied-task benchmark and a search-based question-answering benchmark.
Setup installs vLLM and flash-attention in a dedicated environment, plus the ALFWorld environment with its game files and a pretrained object detector, or a search retriever environment built with FAISS, Hugging Face transformers and datasets, and a retriever service. Training logs can be sent to Weights & Biases.
It is written in Python under the Apache-2.0 license, accompanying a paper accepted at EMNLP 2026.