WorldWeaver gives two generated players persistent shared state and raises its aggregate world score from 81.0 to 105.1 in Minecraft experiments. This morning also brings agent-built physics scenes, a benchmark for future geometry, faster and more controllable video generation, 3D texture repair, community robot data, harness-native agent training, cheaper code reasoning, local coding agents, a medical storytelling game, and Meta AI’s new connected actions.

What Changed Overnight

  • ArXiv’s July 24 announcement batch added new work on shared generated worlds, physical-scene authoring, future-surface reconstruction, video generation, 3D assets, robot data, agent training, coding, and game-based learning.
  • Meta said it began rolling out planning, recurring briefings, research, slide creation, and connected email and calendar actions in the Meta AI app and on meta.ai in selected markets.

Shared Worlds And Physical Scenes

WorldWeaver carries one state across two generated players

Sicheng Mo, Yuheng Li, Ziyang Leng, Krishna Kumar Singh, and Bolei Zhou add persistent world-state registers to a streaming video diffusion model. The registers store shared scene information and each player’s state, then update after every generated chunk instead of asking the model to reconstruct the world from a short window of recent frames.

The project evaluates synchronized two-player Minecraft rollouts. Registers alone raise the aggregate world score from 81.0 to 93.8. Supervising them with agent position, velocity and orientation, a bird’s-eye view, and scene text raises the score to 105.1. The model improves shared building and grounding consistency, but the release does not establish how the approach scales beyond two players or Minecraft scenes.

GS-Agent writes simulation code for a prompted 4D scene

Hongxin Zhang and collaborators divide world construction among agents responsible for assets, material properties, placement, motion, cameras, and lighting. The agents call a physics engine through code, inspect multimodal feedback, and revise the scene.

The project page shows rigid bodies interacting with liquids, sand, cloth, and other deformable materials. Its output remains a simulatable scene with geometry, particles, and code rather than only a rendered video. The initial report does not provide an aggregate task-success rate or authoring-time comparison with human creators.

FutureSurf separates a convincing future view from correct future geometry

Yukun Shi and Minglun Gong train reconstruction methods on the first 75% of a dynamic sequence, then score extracted surfaces on the withheld future. Their dataset contains eight controlled motions, including three falsification cases, with exact meshes for every frame.

For DG-Mesh, the future-to-observed error ratio is 2.7 to 4.1 even on motions recoverable from the observed segment by a fixed rule. Across six animated assets and a second backbone, Deformable-3DGS, the ratio ranges from 2.0 to 6.6. The authors also find that future rendering quality and future-surface accuracy are statistically decoupled, so plausible pixels do not demonstrate correct future geometry.

Video And Asset Control

SANA-Video 2.0 mixes linear and full attention

Junsong Chen and collaborators use linear attention for most layers and insert full-softmax attention at a three-to-one ratio. Residual routes carry the full-attention features into later linear layers.

With 40 sampling steps, the authors report a VBench score of 84.30 in 13.2 seconds at 480p on one H100. Their optimized 5-billion-parameter pipeline produces a five-second 720p video in 13.06 seconds and runs 120 times faster than Wan 2.2-A14B in the reported single-H100 comparison. The paper does not report tests on creator hardware or a game-specific motion set.

GraphVid controls interactions instead of drawing every trajectory

Vedant Shah and collaborators represent subjects and relations as a graph for image-to-video generation. A creator can specify interactions among several objects without drawing a precise track for every subject through overlaps and occlusions.

Against Motion-I2V, GraphVid reports reducing FID by up to 39.9% and FVD by 37.6%. PSNR rises from 9.87 to 15.98 and SSIM from 0.38 to 0.61. The paper also introduces GraphVid-Bench, a dataset with structured interaction annotations; the abstract does not state its size or public-release terms.

Texture++ repairs selected regions of old 3D texture maps

Shuaiwei Wang and collaborators render a 3D asset from several automatically chosen views, enhance masked regions with a diffusion model, then merge the results back into the texture map with a quadtree-based region system.

The method is aimed at low-resolution film and game assets whose UV layouts are fragmented across patches. The authors report better detail and coherence than the tested super-resolution methods, but the published summary gives no aggregate gain, processing time, or release date.

Robot Data And Agent Training

AXIS collects robot demonstrations through a browser

Mengfei Zhao and collaborators let remote contributors teleoperate simulated manipulation tasks in a browser. An automated pipeline checks success, filters trajectories, smooths motion, and adds visual and physics-based variations.

AXIS currently contains more than 50,000 trajectories across 207 tasks. Continual pretraining raises the overall success rate of π0.5 by 5.8% and outperforms a RoboCasa365-pretrained comparison by 37.3%, according to the paper. The largest gains appear under layout, camera, and sensor-noise changes.

OpenForgeRL trains agents inside their actual harnesses

Xiao Yu and collaborators place a proxy between an agent harness and its model calls, record rollouts for reinforcement learning, and run each task in a separate Kubernetes container. The framework covers tool agents, browser agents, computer-use agents, and harnesses including Codex and OpenClaw.

OpenForgeGUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager. OpenForgeClaw reaches 31.7 pass³ and 55.9 pass@3 on ClawEval. The authors say reinforcement learning improves self-checking and tool coverage, while error recovery remains weak.

Error localization reuses the valid prefix of a failed solution

Rajiv Shailesh Chitale and collaborators compare token probabilities with and without task feedback to locate the step where a reasoning trace went wrong. Their method cuts the trace at that point and branches again, keeping the earlier work.

On LiveCodeBench with Qwen3-8B, the method reaches 71.0% pass@64 while generating about 360,400 tokens, compared with 735,000 for independent sampling. The tests use fixed or environment feedback that can identify a failed outcome; performance depends on having that feedback.

Local open-weight coding agents complete up to 87.9% of a data benchmark

Mack Nixon and collaborators evaluate locally deployed coding agents on 20 longitudinal-data preparation tasks that create 102 variables from six study waves. The goal is to keep personal research data off third-party cloud services.

The strongest tested 31-to-35-billion-parameter models reach up to 87.9% average task completion on consumer-grade hardware. The benchmark covers R data cleaning, category harmonization, and multi-wave merging rather than general repository engineering.

Products And Learning

MedGame turns clinical cases into executable stories

Qian Wu and collaborators use one model component to turn a clinical case into states and decisions, then another to create a dependency-aware multimodal plan for an interactive player. MedGame Bench contains 5,000 cases for narrative generation and story direction.

The authors report that task-specific fine-tuning narrows the gap between open and commercial models. Participants in a pilot student study described the system as more engaging and useful than text-only alternatives, but the abstract does not give the participant count or learning-outcome measurements.

Meta AI starts carrying out recurring and connected tasks

Meta says Muse Spark 1.1 now lets Meta AI plan multi-step tasks, connect to email and calendar apps, prepare recurring briefings, research a topic, and create slides. Users can redirect a report or plan while it is being assembled.

Meta says the features began rolling out July 24 in selected markets through the Meta AI app and meta.ai, with WhatsApp and more countries planned later. The announcement does not name the initial markets, supported email and calendar providers, benchmark results, or a complete permission model.

New On The Events Calendar

Wonder News added the ISTE+ASCD Conference 2027 to the calendar. The education-technology conference runs June 27-30 in Boston. Educator program-reviewer applications are open; the organizer says reviewer selection begins August 15 but does not state an application deadline. The call for participation is scheduled for September 1-30, while attendee registration has not been announced.

Watch Next

  • WorldWeaver code, tests beyond two players, and evaluations outside Minecraft.
  • GS-Agent task-success, editing-time, and reproducibility measurements for complete physics scenes.
  • FutureSurf results from methods that explicitly predict geometry beyond the observed window.
  • SANA-Video 2.0 quality and speed tests on creator hardware and longer game-like motion.
  • GraphVid-Bench release details and interaction-control comparisons under heavy occlusion.
  • Texture++ code, processing times, and engine-side checks for seams and material fidelity.
  • Independent AXIS and OpenForgeRL replications, especially under new tasks and recovery failures.
  • MedGame learning outcomes and Meta AI’s supported markets, integrations, and permission controls.
  • ISTE+ASCD’s reviewer deadline, September call for participation, and attendee-registration opening.

This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.