RightSignal · Wednesday, October 21

FocusGUI: helping screen-using AI agents focus on what matters at each step

Fuyao Zhang, Wenwen He, Jiaming Zhang et al. · Submitted 19 Oct 2026 · 199 upvotes yesterday

Preprint · not peer reviewedTrending

Some AI agents operate apps by looking at screenshots, audio, video and their own past actions. The authors add a small adapter that picks the inputs that matter for the next action and passes them to a frozen model, without changing the model itself.

Key finding: Across four models, the adapter raised one accuracy score by 33.2 points and tripled the rate of exact matches compared with the unchanged models.

Why it matters: Agents that pay attention to everything can drift off course. Showing them only what matters at each step makes them more reliable at a low cost.

Read on arXiv

Matrix-Game 4.0: an AI that draws playable game worlds in real time

Jihai Zhang, Runjia Qian, Kaiqin Kong et al. · Submitted 20 Oct 2026 · 177 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

A small game engine keeps track of the world and its rules. A coding agent turns players' ideas into game code. A video model then turns the engine's simple shapes into detailed images.

Key finding: It runs at 480p and 60 frames per second on a single RTX 5090, with 112 milliseconds between a key press and the screen updating.

Why it matters: Separating game rules from the drawing step makes AI-generated worlds consistent and playable on a consumer graphics card. The code and model weights are public.

Read on arXiv

PAR: teaching looped language models to solve much longer problems than they trained on

Pan Wang, Yihao Hu, Guangyuan Dong et al. · Submitted 20 Oct 2026 · 147 upvotes yesterday

Preprint · not peer reviewedTrending

Looped models run the same layers several times to think longer. The authors train each loop to match the state after one more step of the problem, and they keep the model's internal values at a steady scale.

Key finding: A small model trained on puzzles of up to 32 moves reached 95.5% accuracy at 512 moves. Step-by-step reasoning on the same model reached 0.8%.

Why it matters: Giving a model more time to think only helps if it learns to make steady progress. This paper shows one way to teach that.

Read on arXiv

Systems

Specter: running large AI models faster on one GPU with too little memory

Wenfeng Wang, Xiaofeng Hou, Xinfeng Xia et al. · Submitted 20 Oct 2026 · 19 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Mixture-of-experts models are often too large for one GPU, so parts of them, called experts, are kept in ordinary memory and moved over when needed. A small, compressed copy of the model predicts which experts will be needed next, so they can be moved ahead of time.

Key finding: On average it produced each word 1.50 times faster than the best existing system, and up to 1.88 times faster.

Why it matters: It makes large open models more practical on ordinary hardware. The code is public.

Read on arXiv

Your RightSignal

Topics in News

FocusGUI: helping screen-using AI agents focus on what matters at each step

Matrix-Game 4.0: an AI that draws playable game worlds in real time

PAR: teaching looped language models to solve much longer problems than they trained on

Specter: running large AI models faster on one GPU with too little memory

That’s every paper.