ShareVerse: Multi-Agent Consistent Video Generation for Shared World Modeling

A team of researchers has introduced ShareVerse, a video generation framework designed to produce spatiotemporally consistent virtual environments across multiple independent agents, according to a paper posted on arXiv [1]. The framework, detailed in a submission last revised on 3 Jun 2026, addresses a gap in existing world-model research, which has largely focused on single-view or small-scale generation rather than unified shared worlds with multi-agent interaction [1][4]. ShareVerse is built on the CARLA simulation platform and supports 49-frame large-scale video generation, accurately perceiving the position of dynamic agents while maintaining shared world consistency [1][2]. The paper lists Xiaoyun Yuan as the corresponding author [1]. The system integrates three core innovations. First, a dataset for large-scale multi-agent interactive world modeling was constructed using CARLA, featuring diverse scenes, weather conditions, and interactive trajectories with paired multi-view videos — front, rear, left, and right views per agent — along with camera data [2][4]. Second, a spatial concatenation strategy arranges four-view videos from independent agents to model a broader environment and enforce internal multi-view geometric consistency [2][5]. Third, cross-agent attention blocks are embedded into a pretrained video model, enabling interactive transmission of spatial-temporal information across agents. This mechanism guarantees shared world consistency in overlapping regions and reasonable generation in non-overlapping regions [2][4]. To prevent scene drifting over extended sequences, the framework incorporates a spatiotemporal memory retrieval system. Agents dynamically query and fetch historical tokenized scenes from a collaborative pool based on relative spatial proximity, anchoring long-range autoregressive generation without relying on explicit 3D representations [3]. The approach is grounded in multi-agent autonomous driving simulation, where each vehicle operates as a video generation agent [3]. The initial submission on 3 Mar 2026 was 4,242 KB, and the revised version on 3 Jun 2026 expanded to 8,357 KB [1]. The paper is also indexed on Hugging Face, where its abstract and contributions are summarized [5].