Schema: Discovering Unknown Environments via Agentic Program Induction
Schema agent harness lifts ARC-AGI-3 RHAE from 58.7% to 99.2% via program induction.
Schema is an agent harness that learns unfamiliar environments by inducing executable programs instead of recording discoveries in prose. It maintains a persistent program workspace and interfaces for checking programs against interaction history, planning, and executing plans with step-by-step verification. Using the same base model, it raises ARC-AGI-3 RHAE from 58.7% to 99.2%, solves 100% of public DiG-bench games, and reaches the median of the top-50 human MazeBench players. Ablations attribute gains to each harness component.
- Agents store environment theories as executable programs.
- ARC-AGI-3 RHAE rises from 58.7% to 99.2%.
- Schema solves all public DiG-bench games.
- It matches the median of top-50 MazeBench humans.
- Ablations support each harness component.
From the original article
1 min read at arxiv.orgLearning to complete tasks in unfamiliar environments with unknown rules remains a key challenge for LLM agents. Current LLM agents often record their discoveries in prose, which may not provide a compact, explicit account of how the environment works.
By Guanning Zeng · 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.