Wonder reports generating minute-scale video worlds that a user can navigate by moving the camera at 16 frames per second. This morning also brings search-free world-model control, camera-only driving self-play, faster image and video generation, new tests of GUI and repository reasoning, agent-memory systems, reactive robot policies, live-control safeguards, adaptive 3D streaming, and a new tabletop-games calendar entry. Unless noted otherwise, the results are author-reported from new preprints and have not been independently replicated.

What Changed Overnight

  • ArXiv’s July 29 announcement added new work on interactive video worlds, world-model control, visual driving, generation speed, GUI transitions, code reasoning and optimization, agent memory, robot control, and dynamic 3D delivery.
  • Twin Cities Tabletop Games Expo opened exhibitor applications and badge sales for its April 16–18, 2027 event in Brooklyn Center, Minnesota.

Playable Worlds And Generative Media

Wonder reports minute-scale navigable video at 16 FPS

Jiacong Xu and collaborators built Wonder to continue an image or video as an interactive world. A dense camera-coordinate field tells the model how the viewpoint moves, while sparse attention selects relevant tokens from an expanding history instead of attending to the entire sequence. The authors report diverse, minute-scale rollouts at 16 FPS with revisiting of earlier areas and video-conditioned “re-shooting.” The paper does not establish game rules, object interaction, persistent simulation state, or player goals, so “playable” here means camera navigation rather than a complete game loop.

INTACT turns a predicted visual change directly into an action

Junhan Sun, Hao Zhao, and Guofeng Zhang train INTACT on action-labelled trajectories without rewards. On four official LeWM tasks, direct inference reaches 85.78%, 100%, 97.67%, and 97.89% success in 2.9–5.5 milliseconds. Adding local search reaches 96.86% macro success with 384 candidate sequences, compared with 9,000 candidates for pure CEM search. The results are limited to the reported task suite.

Pictura trains driving agents from the camera view they actually receive

Pictura renders every simulated driver’s egocentric image rather than giving the policy hidden vectors for exact poses and velocities. The simulator reaches up to 500,000 agent steps and two million images per second on one H100. Its Alberti policy trains for 50 billion steps, about 35 million simulated kilometres, and transfers without further training to re-rendered Waymo layouts. The abstract reports that it approaches a privileged-vector policy and beats privileged agents on that transfer test, but does not provide road-deployment evidence.

Parallel Decoding Distillation reduces repeated denoising

Neta Shaul, Chao Liu, Arash Vahdat, and Julius Berner train one network evaluation to predict several denoising steps. The method supports a variable inference budget and reports state-of-the-art results at four to eight network evaluations on LTX-2.3 text-to-video/audio, Wan 14B text-to-video, and Qwen-Image. The abstract says video diversity improves over other few-step methods; public side-by-side tests will be needed to judge motion and prompt adherence across creator workloads.

Interfaces, Code, And Agent Memory

Desktop-Delta Bench tests whether agents understand what changed

Desktop-Delta Bench contains 2,013 human-verified cases from Linux trajectories spanning about 15 applications and 50 task domains. Its 463 ordering tasks include 105 with an unrelated decoy frame, while 1,550 before-and-after pairs test five action families and their payloads. Across eight model families, the best exact-match rate is 65.1% without decoys and 65.7% with them. Click F1 reaches 0.96, but drag identification reaches 0.76.

RepoReasoner finds a cross-file reasoning ceiling

RepoReasoner uses traced pytest executions to build output-prediction and call-chain tasks across repositories, then rewrites inputs and outputs to reduce memorization. Seven models were tested. Even with oracle context, the best model reaches 69.1% Pass@1 on output prediction. Call-chain answers tend to have high precision but low recall, and longer context does not consistently help when irrelevant code is present.

Code-optimization rewards need more than a stopwatch

Pierre Chambon and collaborators combine large optimization tests, a calibrated sandbox, and a correctness-plus-speed reward. On DMC-Optim, strict top-50% Pass@1 rises from 18.0% to 31.3% for Qwen 2.5 7B and from 30.7% to 50.4% for CWM 32B. CWM reaches about half the human rate of complexity-class improvements, 14% versus 28%, showing progress without matching the fastest correct submissions.

UniMem decides when experience should become model memory

UniMem routes new or sparse experiences to an episodic buffer and consolidates recurring execution patterns into expandable parameters, without requiring explicit task boundaries. Across long streaming task sequences and three backbone models, the authors report an average gain of 4.0 exact-match points over their comparisons. The abstract does not report the additional training cost or how often consolidation preserves a pattern that later becomes obsolete.

MemLens exposes which stored interactions earn their keep

MemLens treats agent memories as inspectable data objects. Its dashboard estimates each record’s contribution, visualizes hierarchical memory, and compares response quality, retrieval latency, and token use in a study-copilot application. The paper presents an end-to-end management system; evidence that its Shapley-style value estimates generalize across assistants and changing users remains to be shown.

Robot Control, Live Systems, And 3D Delivery

πR² reacts every 40 milliseconds on a real robot

Sungjae Park and Shubham Tulsiani split robot-policy inputs into a fast proprioception channel and a slower vision-language channel, then update actions in one denoising step. Fine-tuned from GR00T-N1.7, πR² replans at about 25 Hz on an A5000—roughly four times the base rate—and reports gains of up to 23% in simulation and 30% on an xArm6 with XHand. Those are maximum task gains rather than one pooled score.

LLM4OSC keeps deterministic code between language and a live show

LLM4OSC has a model propose structured intent against a reviewed device profile; deterministic code then validates addresses, types, ranges, and the final UDP message. On one Max/MSP profile, four backends reach 100% semantic accuracy and 0% wrong sends across the frozen suite only after symbolic processing. The rules backend remains the production default at about 0.05 milliseconds, while language-model backends take roughly three to four seconds.

SplatStream layers dynamic 3D scenes for changing bandwidth

SplatStream divides dynamic Gaussian-splatting scenes into quality and resolution layers, predicts across layers and time, and sends high-importance primitives first. It maps the result to MPEG-DASH sub-representations for adaptive delivery. The abstract describes the codec design but supplies no bitrate, latency, or visual-quality numbers, so its advantage over existing streaming formats cannot yet be quantified from the preprint summary.

Events Calendar

Twin Cities Tabletop Games Expo opens badges and exhibitor applications

The Twin Cities Tabletop Games Expo will run April 16–18, 2027, at the Heritage Center of Brooklyn Center in Minnesota. Exhibitor applications are open for its 13,000-square-foot hall, and attendee badges are on sale; the organizer has not published closing dates for either window. Registration for scheduled game sessions opens Jan. 15.

See the 2027 games and education-technology calendar.

Watch Next

  • Wonder needs public interaction tests that separate camera control and visual recall from simulated objects, rules, and persistent player actions.
  • Pictura’s camera-only self-play can be compared with policies using privileged simulator state on closed-loop safety and rare events.
  • RepoReasoner and Desktop-Delta Bench provide two places to test whether better state tracking improves both code agents and computer-use agents.
  • πR²’s next test is whether its fast proprioception path remains effective when vision changes abruptly rather than merely arriving late.

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