A new generative game engine can render two-player fighting-game interaction at 30 frames per second from two sets of keyboard inputs, its authors report. This edition also covers a rhythm-game chart benchmark, a physics weakness in video diffusion, 3D and XR research, lower-cost coding-agent methods, a mobile agent framework, and concern that fast AI cloning could make games harder to discover. New event calendars list upcoming games, indie-game, and EdTech dates and application windows.

What Changed Overnight

  • WanToFight joined the latest arXiv release with a two-player combat demonstration built on a 1.3-billion-parameter video diffusion model.
  • ChartGenEval introduced corruption-tested feedback for generated rhythm-game charts instead of grading them only against one official note sequence.
  • GamesRadar+ reported Circana analyst Mat Piscatella’s warning that rapid AI cloning could bury original games on crowded digital storefronts.

Playable Generation And Game Evaluation

WanToFight generates a complete two-player match at 30 FPS

WanToFight takes keyboard commands from two players and generates the next frames of a fighting match. The system binds each input stream to a different character, limits each generated video segment to nearby control inputs, and carries cached attention state forward as the match continues.

The authors report 30 FPS at 512×384 on one NVIDIA RTX 5090. The engine uses a four-step distilled version of the Wan-1.3B video model and a pruned decoder. The project is a research demonstration built around The King of Fighters ‘97 footage; the authors do not release an original game or state that the system is licensed for commercial use.

ChartGenEval separates timing, pattern quality, and repetition

Rhythm-game generators can produce many valid note sequences for the same song. ChartGenEval therefore uses an official chart’s timing map without treating its exact notes as the only correct answer.

The evaluation covers seven output axes tested across 80 held-out song groups. In stress tests, its timing analysis recovered artificial offsets of 15, 30, and 60 milliseconds. Rewriting common patterns reduced language-model perplexity by 37%, while deliberately collapsing loops increased self-similarity by 62%. The framework reports separate measurements rather than one overall score.

More denoising steps did not fix long collision chains

The Seriality Gap study asks video models to predict sequences in which one ball hits another and triggers further collisions. Standard bidirectional video diffusion became less accurate as the causal chain grew, even when the researchers added denoising steps.

A single-ball control did not show the same decline, pointing to dependent events rather than video length alone. Autoregressive generation, blockwise generation, and deeper models improved the results more. In a generated game, the same weakness could surface when one hit must change position, damage, score, and the action that follows.

3D Production And XR Interaction

SeamGen learns where artists place UV seams

UV seams decide how a 3D surface is unfolded for texturing, and poor placement can expose visible breaks or create awkward charts. SeamGen learns edge labels from a corpus of existing professional seam layouts instead of relying only on distortion objectives.

Its mesh transformer combines local attention across connected edges with global attention across vertices. The model also supports local repair and user constraints at inference time. The authors report closer alignment with artist-authored layouts than distortion-based and semantic-proxy baselines, though the abstract does not disclose the corpus size or production-time savings.

People distinguished virtual rigidity without haptic hardware

An XR study tested whether people could sense differences in simulated molecular rigidity without physical feedback. Participants directly manipulated real-time C60 molecule simulations in one setting and watched someone else interact in another.

Direct interaction produced an 11.5% just-noticeable difference, compared with 18.5% for observation alone. Participants who interacted first also performed better in the later observation task. The researchers frame the experiment around molecular education and research, showing a measurable perceptual effect from direct interaction even without force-feedback devices.

Developer Agents And Tools

Task-scoped execution cut agent tokens by 91% in a controlled benchmark

The E3 method asks an agent to estimate a task’s scope, try the minimum sufficient path, and expand only when verification fails. On MSE-Bench’s 121 controlled code edits, it matched the strongest baseline’s 100% success while the authors measured 85% lower cost, 91% fewer tokens, and 92% fewer inspected files.

The paper also tested a live GPT-4o agent on an open-source library and found a smaller but similar over-reading pattern. The authors describe the study as a controlled probe, not a measurement of every deployed coding agent.

Line-anchored comments reduced generated code-editing tokens

Another study compared the same requested changes written as a general prompt and as comments tied to exact lines. The line-anchored format reduced generated tokens by 22% for Claude Opus and 58% for Claude Sonnet. Four of seven tested models produced significantly fewer tokens after correction for multiple comparisons.

Correctness rose by two percentage points across the pooled results and by five to seven points for three of five local models. The treatment changed how feedback was delivered, not what change the models were asked to make.

An agent regenerated only the dependency code a repository used

A software-engineering paper tested whether an agent could replace dependencies with locally generated implementations of only the functions a repository exercises. Across 180 repository-dependency pairs, the replacements preserved 99.8% of behavior observed by the study’s baseline checks and reduced exported API surface by 93%.

Those checks do not prove equivalence for unobserved behavior, future updates, or security properties. The authors say the approach is most practical for narrow, stable, well-tested dependencies.

PalmClaw runs an agent loop directly on a phone

PalmClaw is an open-source framework that manages an agent’s sessions, memory, tools, and execution loop on a mobile device. Instead of depending only on sequences of taps and swipes, it exposes device functions as tools with explicit arguments and structured results.

The paper reports an 11.5% relative improvement in task success and a 94.9% reduction in completion time against its strongest baseline. Its tool interface gives mobile apps a way to expose device functions without asking an agent to operate them only through visual taps and swipes.

Storefronts And Discovery

An industry analyst warns that fast clones could bury original games

GamesRadar+ reported on a 404 Media demonstration in which developers used AI coding tools to reproduce existing game ideas in dozens of hours. Circana analyst Mat Piscatella responded that original games could be cloned and buried as release volume rises and players retreat to familiar franchises.

The comments are an industry reaction, not a measured forecast of future storefront volume. Piscatella’s warning is that Steam and console stores could receive more copies while original developers still compete for limited visibility and player attention.

New On The Events Calendar

Wonder News has opened bilingual 2026 and 2027 event calendars with separate application details for organizations and individual attendees. The calendars begin with 18 confirmed games, indie-game, and EdTech events in Korea and abroad, from BIC Festival and G-STAR to gamescom, Tokyo Game Show, Bett, EDUtech Asia, Taipei Game Show, Education Korea, and SXSW EDU.

The nearest dated windows include G-STAR exhibitor applications through July 31, BIC early-bird online tickets through Aug. 5, and Tokyo Game Show one-day ticket sales beginning July 20. For 2027 events, SXSW EDU’s PanelPicker closes July 26, Taipei Game Show’s Indie Game Award closes Aug. 19, and Education Korea’s early exhibitor rate ends Aug. 28. Future newsletters will mention this calendar only when a newly verified event is added; date corrections will remain on the calendar page.

Watch Next

  • Whether WanToFight exposes latency, input accuracy, match consistency, and failure rates beyond its selected demonstrations.
  • Whether ChartGenEval is adopted by chart generators and tested on genres with different note structures.
  • How long video-generated worlds can preserve collision chains, rules, scores, and character state during continuous play.
  • Whether SeamGen publishes its training-data scale and artist study details.
  • Whether task-scoped and line-anchored coding methods retain their savings on large game-engine repositories.
  • How storefronts respond if AI-assisted releases and rapid clones continue to increase.

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