What if you could replace a core part of a Transformer with something simpler and stronger? Researchers from Princeton, NYU, and CMU present Derf. They swapped the standard "normalization" layer with a simple, element-by-element function called Derf (based on a Gaussian error
@jiqizhixin
-
LongVie 2: AI Generates 5-Minute Videos from Single Images
By
–
What if an AI could generate a 5-minute video from a single image, with perfect control and consistency? Enter LongVie 2. It's a "world model" that builds long videos step-by-step. It uses multiple guidance signals (like sketches or text) to control the scene, and special
-
LoPA: 3x Faster AI Inference Without Additional Training
By
–
You could make AI inference 3x faster without any extra training. Researchers from SJTU & Huawei present LoPA just for that! Their new "Lookahead Parallel Decoding" method is a plug-and-play trick. It smartly predicts the best order to generate multiple tokens at once, instead
-
Data-Centric Approach Improves Fake Audio Detection
By
–
What if the key to spotting fake audio isn't a better algorithm, but smarter data? Researchers from SJTU present a data-centric breakthrough. Instead of just building a bigger model, they focus on optimizing which training data to use, prioritizing diversity of sources and AI
-
RoboTracer: Vision-Language Model for Precise Robot Task Execution
By
–
What if a robot could understand "water the flowers from left to right" and precisely measure the height to hover above each one? A team from Beihang University, Peking University, and the Beijing Academy of AI presents RoboTracer. It's a new vision-language model that gives
-
Jacobi Forcing: Generate Multiple Tokens in Parallel Without Quality Loss
By
–
What if you could generate 4-5 tokens at once without losing quality? Researchers from UC San Diego, SJTU, and Snowflake introduce "Jacobi Forcing." It smoothly converts standard autoregressive models into parallel decoders by training them on their own predicted token
-
UniCardio AI Model Cleans Noisy Heart Signal Data
By
–
This AI could fix messy heart data and even generate missing signals. UniCardio is a unified model that uses the relationships between heart signals to clean up noisy wearables, fill in gaps, and translate between signal types. It beats specialized models in denoising &
-
Digital Red Queen: AIs Evolve Through Competitive Self-Play
By
–
Sakana makes AIs "hunt" each other, and they begin to converge in evolution
— 机器之心 JIQIZHIXIN (@jiqizhixin) 10 janvier 2026
Researchers from MIT & Sakana AI present "Digital Red Queen."
They use LLMs in a self-play loop to evolve "warrior" programs that compete in Core War, a classic programming game. Each new warrior must… https://t.co/n9Los852dUSakana makes AIs "hunt" each other, and they begin to converge in evolution Researchers from MIT & Sakana AI present "Digital Red Queen." They use LLMs in a self-play loop to evolve "warrior" programs that compete in Core War, a classic programming game. Each new warrior must
-
JoVA: AI Generates Perfectly Synced Video Speech
By
–
What if an AI could generate a video with perfectly synced speech, not just background noise? Researchers from The University of Hong Kong & ByteDance present JoVA. It uses a single, streamlined transformer where video and audio data directly interact in every layer, avoiding
-
AI Foundation Model Understands 3D Depth in 360° Panoramas
By
–
What if a single AI could understand the 3D depth of any 360° panorama? Insta360 Research & UCSD present Depth Any Panoramas (DAP), a new foundation model. They built a massive, diverse training set and designed a model that learns robust, real-world distance cues without