SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators

What it claims

The one mechanism the existing summary omits: SyncWorld does not strictly require a calibration episode at inference. The abstract states that training on visual calibration contexts teaches the model two things — to interpret actions through visual evidence when calibration is provided, and to leverage interaction history when explicit calibration is unavailable. So there is a fallback path: the model can infer the setup-specific Action–Visual Mapping from accumulated interaction frames rather than a dedicated calibration episode.

The paper also frames the specific failure it targets: under mixed multi-setup training, identical numerical actions map to different visual outcomes across environment, camera view, robot placement, and embodiment, producing "conflicting supervision" and brittle deployment generalization. Calibration-in-context is the mechanism to resolve that conflict at inference rather than at train time.

Limits

This is an abstract-only page — no numbers, no baselines, no benchmark names, no dataset sizes, no compute figures, and no code or model release links appear. Every quantitative claim in the existing summary ("accurate rollouts," "test-time policy improvement") is asserted without a metric or comparison here. An engineer deciding whether to build on this cannot yet judge the magnitude of the effect, which embodiments or environments were tested, or whether weights exist; those must be checked in the PDF. Submitted 8 Sep 2026 (v1).