ASIGAME WORLD

ASIWorldBenchmarking Autonomous Agents
in a Persistent Survival Sandbox

Jianfeng Lin1,2,* Runze Wang1,3,* Haorong Hong1,*
Yuxuan Xie1 Zeno Zeng4 Zhongyuan Liu1,†

1 Lightspeed Studios, Tencent2 HKUST3 USTC4 Hunyuan Team, Tencent

* Equal contribution   ·   † Corresponding author

Play ASIWorld PaperarXiv CodeComing soon
Illustrated agent journeys through a voxel world: Gemini at a workshop, Kimi at a bridge, GLM and DeepSeek in the ruins, and white-clad GPT and orange-clad Claude in the final battle arena. Positions are illustrative, not measured trajectories.

Agent capabilities

Common Claude Code harness · 300-minute main season

6 of 9 agents completed both goals.

Goal attainment, combat cost, and execution strategy at a glance.

Agent capability overview from the manuscript. Outcomes, achievements and actions use the main season; dungeon rates and damage combine three seasons.
Agent01 Goal attainmentMain season02 Dungeon performanceThree seasons03 Action strategyMain season
Goals P / EAchievementsClear rate (%)OuterInnerRoyalKingDamage taken ↓Shadow King
claude-opus-5-5
7/8
81100100100
60
gpt-6-astra
8/8
83100100100
0
claude-opus-5
8/8
8167100100
98
deepseek-v4-flash
7/8
732256100
92
glm-5.3
6/8
6660100100
127
kimi-k3
7/8
586710075
154
gemini-3.8-flash
3/8
616710075
82
gpt-6-sol
3/8
58258367
173
gpt-6-luna
1/8
60––
–
Goal completed · P: production / E: explorationProductionConstructionDungeon combatMaintenanceMovement
How to read this overview

Clear rate: boss defeats per attempt. Damage: mean damage taken in successful Shadow King clears (lower is better; maximum health 100). Three-season observations are not matched repetitions. Action shares describe executed actions, not time; bars are redrawn from the published figure. “–” indicates no attempts or no successful clear.

Overview

ASIWorld evaluates autonomous agents in a persistent survival sandbox. Agents coordinate production and exploration through direct actions or asynchronous programs, while earlier decisions shape later opportunities. Evaluation considers task outcomes, development achievements, survival costs, and execution behavior.

ASIWorld overview: a persistent survival world, an agent harness with asynchronous program execution, and evaluation using execution records and world state.
ASIWorld connects persistent world state, asynchronous agent execution, and evaluation of outcomes, achievements, and costs.

Benchmark Tasks

Two task families test coordinated execution through dependencies, preparation, and changing world state.

Production and exploration task paths, with resource and crafting dependencies, dungeon tiers, and additional attainment requirements.
Task paths and additional attainment criteria in the production and exploration evaluations.

Evaluation

Assessing whether agents achieve their goals, how far they progress, and what attainment costs.

Dungeon performance

Two heatmaps from the manuscript showing clear rate and damage taken per clear for nine configurations across Outer Court, Inner Temple, Royal Tomb, and Shadow King.
Across three seasons: boss defeat rate per attempt (left) and mean damage taken in successful clears (right).

Action composition

Stacked bars from the manuscript showing the share of executed actions devoted to production, construction, dungeon combat, maintenance, and movement in the main season.
Share of executed game actions by category in the main evaluation season.