WorldPlay2 pairs a factorized hybrid control interface with compressed memory and stable distillation, reaching 83.1 average on WBench and 0.105 MEt3R on RevisitBench

It proposes a factorized hybrid control interface that explicitly disentangles low-level movement from high-level semantic interaction and visual content factors: frame-aligned action control handles continuous camera pitch and yaw, discrete longitudinal and lateral movement, camera perspective, and a jumping action, while structured semantic control partitions signals into scene, character, and event fields. Prior work largely used camera poses, discrete keyboard inputs, or pixel-space coordinate fields for viewpoint and character movement, and although language-driven events were introduced, heterogeneous control signals remained hard to represent in a unified way; this work separates controls of different semantic granularity and temporal horizon so the model can learn reusable combinations across them. The paper gives a formula-level definition of the interface and its injection point before the FFN in each Transformer block, and reports ablations: removing structured semantic control raises rotation error from 0.168 to 0.478 and lowers translation accuracy from 98.3 to 88.9, indicating a measurable effect of disentanglement on navigation controllability.

It compresses generated history into compact memory tokens shared by the autoregressive student and the bidirectional teacher, so a long rollout can be partitioned into memory-conditioned clips scored individually instead of jointly processing the entire rollout. Existing methods often retain full-resolution context or rely on retrieval, sparse attention, or explicit 3D representations for long-horizon consistency while treating distillation as an isolated module; this work co-designs memory and distillation so teacher scoring stays tractable over long horizons. The paper reports roughly an order-of-magnitude reduction in sequence length and shows that training on 96 latents yields long-horizon geometric consistency comparable to a full-context baseline (PSNR 18.80 vs 18.47, MEt3R 0.128 vs 0.133) while substantially reducing memory footprint and iteration time.

It proposes Stable Forcing, a long-horizon distillation framework that first uses a PDD-inspired few-step initialization so the autoregressive student produces meaningful few-step rollouts, then uses full-rollout replay to decouple long-horizon rollout from gradient backpropagation, with each chunk performing full few-step sampling and a randomly selected intermediate denoising timestep cached and replayed with gradients. Self Forcing-style methods become unstable over long horizons because error accumulation and few-step sampling drive the student distribution away from the teacher; combining initialization with replay keeps long-horizon distribution-matching distillation stable. Ablations show that removing initialization and full-rollout replay drops WBench average to 73.9 with mode collapse and ground artifacts, replay alone gives 75.2, and the full method gives 83.1; a full-context teacher is competitive (82.3) but costs 261s versus 167s per iteration and hits out-of-memory when extended to 320 latents.

It evaluates systematically on WBench and a self-built RevisitBench: WorldPlay2 averages 83.1, above the previous best baseline Alaya-Evoke-Turbo at 82.0; on RevisitBench it reaches PSNR 19.71, SSIM 0.613, LPIPS 0.318, and MEt3R 0.105, better than five compared models; on interactive-event evaluation it averages 74.7 (EO. 79.2, CI. 70.1). Compared models are limited respectively by memory degradation under fixed context windows, retrieval errors from compounding camera pose drift, metric scale ambiguity across chunks in explicit 3D representations, or temporal flickering; this work maintains long-horizon geometric consistency without error-prone retrieval or sensitive explicit 3D representations. Evaluation covers the navigation split of WBench and RevisitBench with 200 revisit trajectories, and uses a VLM as an automated evaluator over 145 held-out samples for instruction adherence and execution accuracy; the training corpus is 700K navigation clips (30s to 60s) plus 10K interactive event clips (10s to 30s).

The result targets interactive world model settings that need real-time response, multi-turn interactive control, and long-horizon geometric consistency, such as navigation-style exploration and semantic-event-driven environment evolution. The training recipe is a multi-stage curriculum: first train a base video diffusion model for navigation control on spatial navigation data, then integrate the memory compressor following a two-stage regime, then adapt the bidirectional model into a chunk-wise autoregressive model via teacher forcing with PDD few-step initialization, and finally apply distribution matching distillation; on the inference side, computation graph fusion, low-bit quantization, KV caching, and a lightweight VAE reach 16 FPS on 8 H20 GPUs. For a reader, this means the work offers a reusable combination of control interface, compressed memory, and stable distillation that later work can extend to longer horizons, more control types, and larger data scale.

The paper itself notes that characters are still prone to gradual visual and semantic drift and occasionally fail to preserve strict identity consistency during long rollouts, and that scaling to infinite-horizon generation remains an open challenge, with maintaining both infinite rollout stability and long-term geometric consistency without error accumulation unresolved. On evaluation, interactive-event responsiveness relies on a VLM as an automated evaluator whose agreement with human judgment is not elaborated, and RevisitBench consists of 50 10s and 150 30s revisit trajectories, a limited scale. In addition, the 16 FPS real-time claim is tied to 8 H20 GPUs and a specific set of inference optimizations, so behavior on other hardware or without equivalent optimizations is a question to watch.