Paper summary: ABot-World-0: Infinite Interactive World Rollout on a Single Desktop GPU

Problem

Interactive world models must respond to actions in real time over long sessions. Autoregressive video drifts as errors pile up, and most such models need big GPU clusters.

Method

• Data: AAA games, simulation engines and internet videos. An agent (“WorldExplorer”) collects data guided by training feedback, and a pipeline applies 14 deterministic quality checks, VLM assessment, and synchronised action and text annotation.
• Distillation: a bidirectional action-conditioned teacher is progressively distilled into a causal student (teacher forcing, ODE distillation).
• LongForcing aligns long student self-rollouts with an extended-horizon teacher, reducing autoregressive drift.
• Controls: raw keyboard actions, plus reference-character memory for consistent identity.
• Deployment: lightweight VAE decoder, efficient attention, memory-aware scheduling, low-bit DiT inference.

Key results

From the abstract: on a single RTX 5090, it streams 720p at up to 16 FPS, with 1.2 s action-to-first-frame latency and about 19 GiB peak VRAM, and shows competitive controllability on WorldRoamBench and extended rollouts.

Limitations

There’s no dedicated limitations section in the extracted text. The numbers are for optimised low-bit configurations.