Generating an interactive world continuously for over an hour — on a single GPU. A new world model makes this possible. Alaya-EVOKE: From Linear-Scaling Supervision to Endless World Three key innovations that make EVOKE stand apart — a deep dive into each. 🗺 Highlight 1: External Geometric Memory for O(1) Context Traditional world models blow up transformer context as generation grows longer, making sustained generation beyond a few minutes infeasible. EVOKE estimates monocular depth from generated frames, stores them as point clouds in a camera-pose-indexed "World State Bank," and retrieves geometry for the next chunk via pose-based lookup. No matter how long the session runs, each denoiser call receives exactly 1.5 seconds (9 frames) of input. "Extending a session increases only the number of recurrent calls, without increasing the context length of any individual call." 🎓 Highlight 2: Decoupling Supervision Horizon from Gradient Horizon Long-term consistency requires supervision over long horizons — but backprop memory explodes. EVOKE solves this with a 14B Wan2.2 teacher that evaluates full trajectories jointly, while student gradients are computed per-chunk with detached history. A controlled experiment confirms the effect: the long-horizon student (30s supervision) stabilizes at 101% of opening brightness; the short-horizon student settles at 74% (Wilcoxon p=0.016). Long teacher supervision is statistically the key to temporal consistency. ⚡ Highlight 3: CFG-Free 3-Step Inference Tops VBench-2.0 No Classifier-Free Guidance, just 3 denoising steps over a coarse-to-fine latent pyramid — yet EVOKE scores 66.77 on VBench-2.0 (rank 1 of 10 systems), matching multi-step competitors. Generation speed: 2.11 seconds per 1.5-second chunk on a single H200. Decoupling persistent state from denoiser context is the conceptual pivot that finally enables world models to run indefinitely without degrading. #WorldModel# #VideoGenerationAI#

Collaborating with visionaries like Tony Shyu @aineuwave shows exactly what happens when creative intuition meets the power of Topview. Check out this incredible movie he crafted—taking production value to the next level with AI. #AI# #VideoGeneration# #Topview# #CreatorEconomy#

🌍 TL;DR: A training-free way to re-explore a recorded video from new camera angles just dropped. It handles viewpoint changes, hole-filling, and recovering earlier appearances entirely through a frozen model's own attention. Title: World in World: Explore the World with World Models URL: Points 🎥 Four evidence sources (source video, target-view scene, rendered geometry, retrieved history) get converted into clean K/V fed straight into the frozen model 🧭 CGAR routes queries toward reliable source-video tokens using persistent point correspondence and camera geometry 🎚 EWA independently tunes each auxiliary channel's contribution using existing attention outputs, with zero extra network evaluations 🏆 Top VBench score of 85.192 and lowest rotation error (2.8326°) against 6 baselines including ReCamMaster 🔧 Same framework also powers bullet-time rendering, video stabilization, editing, and motion transfer Getting this much versatility with zero retraining feels like a win for evidence engineering over model adaptation. #WorldModels# #VideoGeneration#

Hide an object behind a wall in a video generation model, and it might come back as something else entirely. This paper tackles that object permanence failure head-on. Title: Training Object Permanence in World Models URL: 📝 Overview The paper trains video models on object permanence and solidity — cognitive abilities human infants develop by six months old. It builds WROP, a benchmark of 150 synthetic task generators producing 1.5M training samples, and PWM-WROP, a 16B-parameter model fine-tuned on it. ❓ Problem Solved Models like Sora let objects vanish behind occluders and reappear as different objects, or pass straight through solid barriers. These failures undermine higher-level reasoning about collisions and cause and effect. 💡 Method & Approach 150 Blender-generated tasks are organized into six occlusion and solidity families. Structural parameters like object count and trajectory vary systematically while surface parameters like color and lighting are randomized, preventing models from succeeding through memorization. PWM-WROP fine-tunes NVIDIA's Cosmos3-Nano on this data. 📊 Results In 361 blind human pairwise comparisons, PWM-WROP ranked first among true-continuation models (Elo 1679.5), 224 points ahead of the runner-up. It topped every model on static occlusion tasks, but still struggled on solidity tasks like collisions. 🌍 Use Cases The team released the training data, model weights, and PWM, a native-PyTorch training stack for AWS Trainium2 — laying groundwork for physically grounded world models. #WorldModels# #VideoGeneration#

Can video generation models actually reason about the visual world — or do they just make things look plausible? These sound similar but are fundamentally different. Placing objects correctly, respecting physical laws, and solving multi-step tasks with coherent intermediate states: this is what "Visual Grounded Intelligence" means. VGI-Bench quantifies how far current models are from achieving it. Across 27 tasks and 810 instances, the strongest model — Seedance 2.0 — reaches only 51.0% overall. Three failure modes appear repeatedly: physical collapse (unrealistic deformations and object penetrations), rule violations (reaching plausible end states while ignoring constraints), and object inconsistency (losing temporal identity across frames). Open-source models lag far behind commercial ones, with the strongest open-source model at 19.1%. What the denoising analysis reveals is even more telling. Self-correction during intermediate generation steps occurs in fewer than 1% of transitions. Wrong-to-wrong transitions at mid-denoising reach 23.1%, and later denoising steps primarily refine early hypotheses rather than fixing errors. Fine-tuning on 1M synthetic samples improved planning and spatial skills but degraded temporal and physical capabilities — a tradeoff that points to deeper architectural constraints. VGI-Bench makes clear that visual reasoning remains a genuine unsolved frontier for video generation. Title: VGI-Bench: Probing Visual Intelligence in Video Generation Models URL: #VideoGeneration# #VisualReasoning#

A video generation model can produce plausible clips — but can it reproduce the correct distribution of futures? These are very different things. Title: PAWBench: How Far Are We from Probabilistically Aligned World Modeling? URL: 🎯 Overview Just as a coin should land heads roughly 50% of the time over many flips, a video generator should reproduce physically correct outcome frequencies — not just look realistic. PAWBench is the first systematic benchmark for this "probabilistic alignment," evaluating 50 scenarios across 11 video generation models. 🔍 The Problem Prior evaluations focus on perceptual quality (FID, FVD) or diversity, but miss whether models generate each outcome at the right frequency (calibration) and whether they can produce all possible outcomes at all (coverage). PAWBench fills this gap. ⚙️ Evaluation Method PAWEval generates K=50 video rollouts, maps them to terminal outcomes via Gemini 3.5 Flash, and scores them using total variation distance (TVD) for calibration and valid-support recovery rate for coverage. Eight physical mechanism groups are tested: tossing, rotation, collision, material transition, and more. 📊 Key Results The best model (Cosmos 3 Super I2V) achieves TVD of only 20.5 (ideal = 0). The average TVD across all models is 31.2 — far above the 8.33 expected from finite-sample chance effects alone. Not a single model simultaneously achieves accurate probabilities, broad coverage, and high scoring reliability. ⚡ What Interventions Reveal Three approaches were tested: prompt engineering, coupled noise sampling, and LoRA fine-tuning. Each improved only partial metrics. Models consistently underreact to physically causal interventions and overreact to non-causal visual or textual cues. For research applying video generation to robotics, autonomous driving, or physical simulation, probabilistic alignment is now an essential evaluation axis. #VideoGeneration# #WorldModels#

🎬 "The video looks realistic — but did it actually complete the task?" Video generation evaluation finally goes outcome-oriented. SemComp-Bench: Benchmarking Semantic Task Completion in Video Generation 💡 Overview When generating videos that complete tasks like "cook a dish" or "prune a plant" from reference images, current models reproduce procedures reasonably well — but systematically fail to be evaluated on whether the intended end-state was actually achieved. SemComp-Bench introduces a 6-domain, 1,273-instance dataset and a dual-axis evaluation framework powered by Doubao-Seed-1.8 VLM to close this gap. ⚠️ The Problem Existing benchmarks emphasize appearance consistency and intermediate procedural steps, leaving "final outcome achievement" and "semantic grounding in reference images" unevaluated as a joint criterion. 🔬 Evaluation Framework: Two Independent Dimensions · OA (Outcome Achievement) Score: ALL four criteria must pass — outcome realization, semantic grounding, entity consistency, and global visual continuity · GR (Generation Reliability) Score: average across five failure-oriented criteria — physical plausibility, visual clarity, artifact-free rendering, spatiotemporal coherence, text integrity 📊 Results (Detailed Instruction Condition) · OA leader: HunyuanVideo-1.5-720P at 37.8% — the best model still falls short of 40% · GR leader: Seedance 2.0 at 91.8% — yet ranks 4th in OA at just 20.0% · T2V (text-only) OA: only 0.6–5.0%, confirming visual reference is indispensable · Biggest bottleneck: Within-Scene Spatiotemporal Coherence (0.328–0.739) → OA and GR are independent capabilities. Generating polished videos and completing tasks are entirely different problems. #VideoGeneration# #Benchmark#

Are we actually evaluating world models — or just collecting scores we can't explain? That gap is exactly what this research tackles. HarnessEval-W: Agentifying the Evaluation of Visual Worlds ❓ What's broken about current world model evaluation? 💡 Existing benchmarks return scalar scores with no verifiable reasoning chains. There's no automated way to verify physics correctness, causality, or state persistence the way a human would — making it impossible to know why a model scored what it did. ❓ How does HarnessEval-W actually work? 💡 It uses a three-level agent hierarchy. Cases are routed to specialized skills, sub-agents answer specific sub-questions, and parent agents aggregate evidence into a final score — with every step recorded in an auditable evidence tree. Evaluation spans 8 settings across 3 categories (observation quality, transition correctness, world persistence), covering 330 cases and 18 models. ❓ How does it compare to existing benchmarks? 💡 Physical transition pairwise accuracy improves from WBench's 31.9% to 71.7%, and the draw rate collapses from 52.2% to just 1.8%. Human alignment is strong: Spearman ρ=0.93 for intentional transitions and ρ=0.87 for physical transitions, validated against 5,000 pairwise A/B human judgments. ❓ What does it reveal about today's models? 💡 Seedance 2.0 leads at 75.5, followed by Wan 2.7 (75.0) and Kling 3.0 (74.4). Fine-tuning analysis exposes sharp capability trade-offs: DreamX-World gains +4.8 on exploratory transitions but loses −11.9 on intentional ones. And render quality vs. physical observation correlates at r=−0.04 — nearly independent capabilities. #WorldModel# #VideoGeneration#

🎬 Subject-driven T2V that keeps a reference subject's identity even as it shuttles across domains—real ⇄ fantasy. Title: DomainShuttle: Freeform Open Domain Subject-driven Text-to-video Generation URL: DomainShuttle reconciles subject fidelity with flexible style adaptation. Three highlights worth your attention. 🧬 Domain-MoT Video and reference image are processed in two independent branches; the reference branch uses Domain-aware AdaLN, modulated by time plus a domain attribute (real human / object / background / fantasy subject). Text cross-attention is frozen to preserve the base model's language guidance. 📐 Video-Reference DualRoPE Reference tokens get a separate RoPE space from video tokens for precise subject-level spatial control. Video starts its temporal index at 1, reference is fixed at 0, and multiple subjects (or multiple images of one subject) are organized via positional offsets. 🔗 Cross-Pair Consistent Loss Training uses two different reference sets at the same timestep, suppressing overfitting to single-frame redundancy and extracting the subject's intrinsic features—independent of irrelevant visual properties. Cross-domain subject consistency hits CD-Score 0.861, +18.7% over SOTA (Kling 1.6 is 0.725). A practical win for real⇄fantasy style transfer. #VideoGeneration# #GenerativeAI#

🎬 Distilled autoregressive video models are fast but tend to drift from human preferences. Astrolabe answers that challenge by doing RL alignment in the forward process, with no re-distillation and no reverse-process unrolling. Title: Astrolabe: Steering Forward-Process Reinforcement Learning URL: 📝 Overview Astrolabe is a reinforcement learning framework that aligns distilled autoregressive (AR) video models with human visual preferences. Its defining feature is doing RL in the forward process rather than via conventional reverse-process optimization. It is a large 53-page, 37-figure study. ❓ Challenges Solved Distilled AR video models suit efficient streaming generation but tend to misalign with human preferences. Worse, existing RL doesn't fit these architectures naturally: it typically needs either expensive re-distillation or solver-coupled reverse-process optimization, both heavy and hard to scale. 💡 Methodology & Proposed Approach It rests on three innovations. ・Negative-aware fine-tuning contrasts positive and negative samples at inference endpoints to establish an implicit policy-improvement direction without unrolling the reverse process ・A streaming training scheme generates sequences progressively via a rolling KV-cache, applying RL updates only to local clip windows while keeping long-range coherence through prior-context conditioning ・A multi-reward objective integrates uncertainty-aware selective regularization and dynamic reference updates to mitigate reward hacking, the collapse where only the apparent score rises 🎯 Use Cases It fits real-time streaming video generation where you want to align an efficient distilled model with preferences while preserving its speed. It applies across multiple distilled AR video models and raises quality without sacrificing inference efficiency. 📊 Significance and Results ・By avoiding the heavy paths of re-distillation and reverse-process unrolling, it addresses computational efficiency bottlenecks ・Combining forward-process negative awareness, streaming updates, and reward-hacking mitigation, it provides a robust, scalable alignment solution ・It demonstrates effectiveness across several distilled AR models, with detailed quantitative evaluation and ablations #VideoGeneration# #ReinforcementLearning#