# Video World Model

The work introduces Honeycomb, a video world model built on HexMemory, a low-rank representation that stores scene features in a fixed-size memory of three spatial plus three spatiotemporal planes; a feed-forward writer maps each generated chunk into new plane features, warps the previous planes while preserving their dimensions, and fuses them via confidence-weighted pooling plus a learned residual correction, while a reader retrieves latents to condition subsequent generation; on WorldScore and RealEstate10K it reports stronger generation quality and revisit consistency with feature storage held constant throughout generation.

The work introduces Honeycomb, a video world model built on HexMemory, a low-rank representation that stores scene features in a fixed-size memory of three spatial plus three spatiotemporal planes; a feed-forward writer maps each generated chunk into new plane features, warps the previous planes while preserving their dimensions, and fuses them via confidence-weighted pooling plus a learned residual correction, while a reader retrieves latents to condition subsequent generation; on WorldScore and RealEstate10K it reports stronger generation quality and revisit consistency with feature storage held constant throughout generation.

LOCI introduces a hybrid spatial-memory architecture in which 15 of 30 transformer blocks keep a key-value cache of past observations while the other 15 are restricted to the current chunk and complemented by a recurrent linear-attention memory whose reads and writes are conditioned on projective camera geometry, with recurrent readouts flowing into subsequent cache-backed blocks to supply their queries with accumulated scene context; on the public MIND memory benchmark and on held-out recorded trajectories it reproduces revisited content more faithfully than representative world models and a same-recipe full-softmax model, lowers peak memory at equal length by about 30% with full history, and streams long videos at constant memory with a bounded bank of retained observations.

LOCI introduces a hybrid spatial-memory architecture in which 15 of 30 transformer blocks keep a key-value cache of past observations while the other 15 are restricted to the current chunk and complemented by a recurrent linear-attention memory whose reads and writes are conditioned on projective camera geometry, with recurrent readouts flowing into subsequent cache-backed blocks to supply their queries with accumulated scene context; on the public MIND memory benchmark and on held-out recorded trajectories it reproduces revisited content more faithfully than representative world models and a same-recipe full-softmax model, lowers peak memory at equal length by about 30% with full history, and streams long videos at constant memory with a bounded bank of retained observations.

Memorizon proposes a training recipe in which each scored chunk retrieves its own top-k latents by camera co-visibility and their union forms a bounded shared memory bank, so a training sample can cover a span of any length while the attention sequence stays bounded; revisit consistency improves on every split over a sliding-window baseline, and a span long enough to reach the first visit adds further gains at some cost in image quality.

Memorizon proposes a training recipe in which each scored chunk retrieves its own top-k latents by camera co-visibility and their union forms a bounded shared memory bank, so a training sample can cover a span of any length while the attention sequence stays bounded; revisit consistency improves on every split over a sliding-window baseline, and a span long enough to reach the first visit adds further gains at some cost in image quality.

Physis-Lang treats physical language as a shared, optimizable representation across data curation, model training and video generation, using a physics-aware critic and PhysCapBench to drive an agentic loop that refines physical captioning instructions and language-guided retrieval to cover missing physical processes, yielding consistent gains on four physical video benchmarks with Wan and Cosmos backbones, where models built on the open-source Cosmos3-Nano surpass Veo 3.1 on Physics-IQ Verified, PhyGround and PhyGenBench.

Physis-Lang treats physical language as a shared, optimizable representation across data curation, model training and video generation, using a physics-aware critic and PhysCapBench to drive an agentic loop that refines physical captioning instructions and language-guided retrieval to cover missing physical processes, yielding consistent gains on four physical video benchmarks with Wan and Cosmos backbones, where models built on the open-source Cosmos3-Nano surpass Veo 3.1 on Physics-IQ Verified, PhyGround and PhyGenBench.

The work proposes WorldAttention, an attention architecture for interactive video world models that combines a Hierarchical KV Cache (HKV) storing and retrieving page-level KV across GPU, CPU, and NVMe tiers with a Hybrid Sparse Attention (HSA) that fuses a linear global branch and a head-adaptive block-sparse branch, reporting subject consistency of 0.9472 on VBench-Long and 0.9668 on InterVBench, a 14.02x sparse-kernel speedup over FlashAttention-3, a 2.21x end-to-end speedup, and 22.0 FPS on a single NVIDIA H100.

The work proposes WorldAttention, an attention architecture for interactive video world models that combines a Hierarchical KV Cache (HKV) storing and retrieving page-level KV across GPU, CPU, and NVMe tiers with a Hybrid Sparse Attention (HSA) that fuses a linear global branch and a head-adaptive block-sparse branch, reporting subject consistency of 0.9472 on VBench-Long and 0.9668 on InterVBench, a 14.02x sparse-kernel speedup over FlashAttention-3, a 2.21x end-to-end speedup, and 22.0 FPS on a single NVIDIA H100.

The work presents WorldCrafter, a camera-controllable autoregressive video world model that learns a camera-queryable implicit 3D-aware memory, compressing historical latent frames into a fixed budget of memory tokens read out under the requested viewpoint and injected into the video diffusion transformer before denoising, so that a single image or text prompt supports streaming, minute-scale scene exploration with better revisit consistency and camera-control accuracy than the evaluated baselines while preserving visual quality.

The work presents WorldCrafter, a camera-controllable autoregressive video world model that learns a camera-queryable implicit 3D-aware memory, compressing historical latent frames into a fixed budget of memory tokens read out under the requested viewpoint and injected into the video diffusion transformer before denoising, so that a single image or text prompt supports streaming, minute-scale scene exploration with better revisit consistency and camera-control accuracy than the evaluated baselines while preserving visual quality.