ReactiveGWM: Steering NPC in Reactive Game World Models

Configuration Overview

A clear comparison across three factors:
training source, strategy data usage, and whether the NPC follows strategy commands.

Vanilla

  • Training source Trained on this game
    (no strategy labels)
  • Strategy data — Not used
  • NPC strategy following ✗ Does not follow strategy commands

ReactiveGWM base

  • Training source Trained on this game
    (with strategy labels)
  • Strategy data ✓ Used
  • NPC strategy following ✓ Follows strategy commands

ReactiveGWM transfer

  • Training source No training on this game
    — strategy reused from a different game
  • Strategy data — Not used
  • NPC strategy following ✓ Follows strategy commands
    (zero-shot on this game)

Control Interface Overview

The left side shows the player issuing actions via buttons.
The right side shows the NPC receiving high-level guidance via strategy.

Player Side

Player side controlled by buttons

Left: The player directly controls actions through button inputs.

NPC Side

NPC side controlled by strategy

Right: The NPC follows high-level intent driven by strategy.

Street Fighter II Comparison

SF2 Button Mapping

XLight Punch (LP)
YMedium Punch (MP)
ZHeavy Punch (HP)
ALight Kick (LK)
BMedium Kick (MK)
CHeavy Kick (HK)

Prompt Detail

💡 Click the corresponding video to view its prompt details.

No video selected yet.

Street Fighter III Comparison

SF3 Button Mapping

XHeavy Punch (HP)
YMedium Punch (MP)
ZHeavy Kick (HK)
ALight Punch (LP)
BLight Kick (LK)
CMedium Kick (MK)

Prompt Detail

💡 Click the corresponding video to view its prompt details.

No video selected yet.