Interactive video world models share an awkward problem: users must control three things at once — the viewing perspective, the character's actions, and the camera. Mainstream practice hides view modes inside text captions, conflates action with camera motion, or wires a separate conditioning pathway for each control, so the three channels never see each other and conflicts get ground out by data alone. On October 23, Alibaba's Token Hub team published World-in-Context (WiC) on arXiv with the opposite bet: stop building separate roads, turn all three controls into explicit tokens, and put them in one shared self-attention space. Paper: https://arxiv.org/abs/2610.33169.
Three knobs, one panel
The core is In-Context Conditioning (ICC): view, action, and camera controls no longer travel dedicated pathways but join visual latents as explicit tokens in the same self-attention space. Two findings stand out from the team's analysis. First, explicit view tokens disentangle conflicting dynamics and unlock emergent view switching from single-view training data. Second, action and camera inputs are complementary — action drives local responsiveness while camera governs global stability, and neither substitutes for the other. To preserve temporal continuity across execution intervals, the paper introduces Interaction-Transport RoPE (IT-RoPE), which parameterizes temporal attention with accumulated control transitions.
5B matches 22B
The numbers are the selling point. Built on the Wan2.2-TI2V-5B backbone, WiC scores an 81.5 navigation average on WBench, which the paper reports as on par with the 22B EchoWM (81.6) — the omni world model JD open-sourced in late August. A 5B model reaching 22B territory crosses more than four times the parameter count. The gap widens on perspective switching: 93.8% success versus 18.8% for existing 5B baselines, a 75-point spread. One caveat the project page states explicitly: the 93.8 figure comes after perspective-switching fine-tuning. Training stays lightweight — both stages optimize only LoRA, the condition encoder, and IT-RoPE, leaving the backbone untouched. The GitHub README also lists planned inference defaults of 832×480 resolution, 81 frames, and 16 FPS.
Cold water, and so what
The cold shower: the repository TODO shows inference code, training code, and WiC weights all unreleased, officially "under review." Every headline number is self-reported, third-party replication is impossible for now, and developers who want to verify can only watch the demo videos on the project page. Still, the direction is worth remembering: making control interfaces explicit tokens mirrors what LLMs did with tool calls and reasoning traces — once control signals live in the context, the model can genuinely attribute why the camera should cut now. Over the past two months the world-model track has accumulated narratives about world memory, engine-derived rewards, and evaluation surveys; WiC supplies the "control bus" piece. Whether 5B truly stands next to 22B waits for released weights and independent benchmarks. Until then, it has already shown one thing: the payoff from unifying interfaces may beat stacking parameters.