What they did

The authors adapted the 33B MiniMax-H3 video generator without adding a dedicated action module. Each action is represented as structured natural-language instructions for character behavior and camera movement, then aligned with the corresponding temporal video latents.

They introduced temporal attention routing, which limits each instruction’s influence to its intended interval and reduces leakage between successive actions. Training used 8,000 gameplay samples, 10,000 LoRA optimization steps, and 0.199% trainable parameters.

Key findings

• The unmodified MiniMax-H3 already exhibited zero-shot responses to natural-language instructions for character behavior and camera motion.
• Temporal attention routing made control more precisely localized in time by restricting instructions to designated video intervals.
• Effective joint character and camera control required updating only 0.199% of the 33B-parameter generator.
• The adapted model preserved strong video-generation quality and generalized to scenarios not seen during training, according to the authors.

Why it matters

The work suggests that capabilities learned during large-scale video pretraining can double as a control interface. Rather than building a separate action system, developers may be able to turn sufficiently capable video generators into interactive environments through lightweight alignment of language instructions with time.

Caveats

The reported evidence centers on a single base generator and a relatively small gameplay dataset, so it is unclear whether the method transfers to other video architectures or more complex interactive tasks. The abstract also does not provide quantitative control, generation-quality, latency, or long-horizon consistency results, leaving the practical limits of real-time world control unresolved.