ASIWorld: Benchmarking Autonomous Agents in a Persistent Survival Sandbox
We introduce ASIWorld, a survival sandbox that combines persistent world and character state with character-scoped observations and a programmable asynchronous interface to study sustained goal execution.
ProofPaper ↗
Key points
- Evaluating autonomous agents requires examining whether they can translate goals into coherent action across extended interactions.
- In persistent environments, this requires coordinating interdependent actions under incomplete observations, accounting for delayed effects, and adapting as earlier decisions shape subsequent opportunities.
- Agents pursue interdependent production and exploration tasks through direct actions or agent-authored programs that continue running between model calls.
- Analyses of execution trajectories and performance across difficulty levels reveal different preparation, combat, and recovery behaviors among configurations with similar task outcomes.
Sources (1)
- [1]ASIWorld: Benchmarking Autonomous Agents in a Persistent Survival SandboxarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 10, 10:59 AM
We introduce ASIWorld, a survival sandbox that combines persistent world and character state with character-scoped observations and a programmable asynchronous interface to study sustained goal execution.
Evaluating autonomous agents requires examining whether they can translate goals into coherent action across extended interactions.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 10, 2026Anthropic is cutting off its internal evaluations from the internet
- Oct 10, 2026What Really Works for Universal Multimodal Embeddings? A Controlled Study of Contrastive Training Strategies
- Oct 9, 2026ICYMI: What landed for AI builders in September 2026
- Oct 8, 2026[AINews] not much happened today
- Oct 8, 2026Google brings agentic AI to Gemini, starting with businesses
- Oct 8, 2026Show HN: Let your AI agents paint big arrows, boxes and text on your screen