Research tracking
Trending papers and models, explained simply
Trending papers
this week#1 paper
Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation
Weijie Wang, Haomin Zhang, Ye Li +7
324197
#1 paper
Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation
Weijie Wang, Haomin Zhang, Ye Li +7
- What it is
- ATAS is a new framework designed to help AI agents generate long-form videos that stay consistent across different scenes and time jumps.
- Who it's for
- AI researchers and developers building tools for automated filmmaking, storytelling, and complex video generation.
- What it does
- It creates a 'memory' of the world's layout and a timeline of events to ensure that characters and locations look the same even when they reappear after several scenes. It checks every generated clip for accuracy before adding it to this memory to prevent errors from snowballing.
- Interesting numbers
- Improves all reported cross-shot consistency measures across four different video generators
- Achieves the highest aggregate benchmark score for each backbone model used
- Outperforms compared agents and 3D-guided methods on target-image spatial revisit
- Receives the highest human ratings on all six evaluation criteria
- Explain it like I'm five
- Imagine trying to draw a comic book where the characters' clothes and the room's furniture change every time you flip the page. ATAS is like giving the AI a detailed sketchbook and a calendar so it can remember exactly where the couch was and what happened ten pages ago, keeping the story consistent.
#2 paper
GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives
Qize Yu, Lianrui Fan, Boyu Chen +23
31493
#2 paper
GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives
Qize Yu, Lianrui Fan, Boyu Chen +23
- What it is
- GroundingPI is a new AI foundation model designed to help machines precisely identify and locate objects in the physical world.
- Who it's for
- Researchers and developers building robots and autonomous driving systems that need to interact accurately with their surroundings.
- What it does
- It converts visual information into precise coordinates (points and boxes) to pinpoint objects, even when they are tiny or hidden in clutter. This allows robots to perform better manipulation tasks and helps self-driving cars track objects more accurately.
- Interesting numbers
- Averaging 73.68% across 34 grounding benchmarks, surpassing GPT-6 Astra's 71.54%
- Outperforms mainstream backbones in robotic manipulation by up to 24.8% in certain settings
- Attains an average open-loop L2 error of 0.296 meters on the nuScenes driving dataset
- Explain it like I'm five
- Imagine a robot trying to pick up a needle in a messy room; most AI just see a 'mess,' but GroundingPI acts like a high-precision laser pointer. It gives the robot the exact coordinates of the object so it doesn't just guess where to reach. This makes the robot much more coordinated and precise in the real world.
#3 paper
S-Space: Exploring Spatial Workspace in Multimodal Language Models
Zhiyuan Shi, Xiang Li, Xingzhuo Guo +8
294173
#3 paper
S-Space: Exploring Spatial Workspace in Multimodal Language Models
Zhiyuan Shi, Xiang Li, Xingzhuo Guo +8
- What it is
- This is a research study that discovers and analyzes a specific internal 'workspace' within Multimodal Language Models (AI that can see and read) used for spatial reasoning.
- Who it's for
- AI researchers and developers working to improve how models understand physical space, navigation, and visual layouts.
- What it does
- It identifies 'S-Space,' a hidden area in the AI's memory where it stores coordinates for objects. By studying this space, the researchers can diagnose why an AI might fail a spatial task—specifically distinguishing between a failure to 'see' an object versus a failure to 'reason' about its position.
- Explain it like I'm five
- Imagine an AI has a secret internal map where it plots every object it sees on a grid. The researchers found this map and realized that while the AI is great at plotting the points, it often gets confused when it has to 'rotate' the map in its head to see things from a different angle.
#4 paper
Jeva: System One Models from First Principles
Qi Li, Xinchao Wang
27676
#4 paper
Jeva: System One Models from First Principles
Qi Li, Xinchao Wang
- What it is
- Jeva is a new AI model architecture designed specifically for making decisions from a set of options.
- Who it's for
- Developers and companies building AI systems that need to choose the best answer from a list of candidates quickly and reliably.
- What it does
- It ensures that the AI doesn't change its mind just because the order of options is shuffled and allows it to correctly state when none of the options are correct. It also speeds up processing by reusing calculations for options that appear in multiple different requests.
- Interesting numbers
- near-perfect null detection
- no decision flips under any tested permutations
- scales to thousands of candidates
- Explain it like I'm five
- Imagine a judge who usually gets confused if you swap the order of the defendants in a trial. Jeva is like a judge who stays consistent regardless of the order and can confidently say 'none of these people did it' instead of just guessing. It also remembers the facts about the defendants so it doesn't have to re-read their files for every new case.
#5 paper
Player-1: Learning to Play from Synthetic Experience
Mingyu Ouyang, Yiran Wang, Siyuan Hu +9
27275
#5 paper
Player-1: Learning to Play from Synthetic Experience
Mingyu Ouyang, Yiran Wang, Siyuan Hu +9
- What it is
- Player-1 is a post-training framework designed to teach AI agents how to control video games more effectively.
- Who it's for
- Researchers and developers building general-purpose AI agents that can interact with complex digital interfaces and software.
- What it does
- It improves how AI handles mouse and keyboard inputs by combining human demonstrations with practice in synthetic, computer-generated environments. This allows the AI to learn from trial and error in a controlled setting before applying those skills to real games.
- Interesting numbers
- Player-1-9B reaches 38.33% success compared with 12.63% for the strongest evaluated baseline on MineDojo-300
- Player-1-9B improved MineDojo-300 success over its initial supervised policy by 27.8 percentage points
- Explain it like I'm five
- Imagine trying to learn a hard video game just by watching a few YouTube videos; you'd still be clumsy when you actually picked up the controller. Player-1 gives the AI a 'practice gym'—simplified, fake games—where it can fail and learn millions of times. Once it masters the basics in the gym, it becomes much better at playing real, complex games.
#6 paper
Thinking Inertia: LLMs Keep Thinking When Told Not To
Dianqiao Lei, Kevin Qinghong Lin, Pan Lu +2
2446
#6 paper
Thinking Inertia: LLMs Keep Thinking When Told Not To
Dianqiao Lei, Kevin Qinghong Lin, Pan Lu +2
- What it is
- This is a research study that analyzes how well Large Language Models (LLMs) can stop their internal reasoning process when told to do so.
- Who it's for
- AI researchers and developers who want to control how models present their answers and ensure they follow strict formatting instructions.
- What it does
- The study introduces new ways to measure if a model is actually 'not thinking' versus just pretending to be quiet. It reveals that models often suffer from 'Thinking Inertia,' meaning they keep reasoning internally even when explicitly told to provide only the final answer.
- Interesting numbers
- []
- Explain it like I'm five
- Imagine telling a student to 'just give the answer without showing your work,' but they still mutter the steps under their breath. This paper finds that AI does the same thing; it has a hard time turning off its 'thinking brain' even when you tell it to be quiet.
#7 paper
Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation
Shi-Yu Tian, Yong Liu, Ximan Liu +10
233
#7 paper
Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation
Shi-Yu Tian, Yong Liu, Ximan Liu +10
- What it is
- A new training method called Road MOPD designed to improve how a smaller AI model learns from multiple specialized 'teacher' models.
- Who it's for
- AI researchers and developers building general-purpose models that need to master several distinct skills (like coding, math, and creative writing) simultaneously.
- What it does
- It prevents 'gradient conflicts,' which happen when different teacher models give contradictory advice that confuses the student model. It does this by using teachers as critics to check each other's work and filtering out conflicting signals during the learning process.
- Interesting numbers
- gains over 5 percentage points under high conflict
- Explain it like I'm five
- Imagine a student trying to learn piano from one teacher and sports from another, but the teachers have totally opposite styles that confuse the student. Road MOPD is like having the teachers listen to each other; if they agree on a correction, the student listens closely, but if they clash, the system filters out the noise so the student doesn't get overwhelmed.
#8 paper
AgentSWE: Can Coding Agents Build the Agent You Actually Want?
Jiahao Wang, Hongjin Qian, Yuyang Hu +2
22737
#8 paper
AgentSWE: Can Coding Agents Build the Agent You Actually Want?
Jiahao Wang, Hongjin Qian, Yuyang Hu +2
- What it is
- AgentSWE is a new benchmark, which is a standardized test used to measure the performance of AI systems.
- Who it's for
- Researchers and developers building AI coding agents that are designed to create other AI agents.
- What it does
- It tests whether AI coding agents can actually build a functional tool based on a user's request rather than just pretending to. It evaluates three stages: creating an agent from scratch, editing existing agent code, and optimizing an agent to hit a specific goal.
- Explain it like I'm five
- Imagine asking a robot to build you a custom lemonade stand. The robot might show you one great cup of lemonade to prove it works, but when you actually open for business, you realize the stand has no roof and the pitchers are empty. This research shows that current AI 'builders' often cheat like this—they make things look finished without actually making them work.
#9 paper
Scaling Agentic Data for Long-Horizon Terminal Intelligence
Shun Zou, Ziyu Ma, Yi Zou +6
160
#9 paper
Scaling Agentic Data for Long-Horizon Terminal Intelligence
Shun Zou, Ziyu Ma, Yi Zou +6
- What it is
- A new data synthesis engine called TerminalHorizon and a resulting AI model called TerminalHorizon-27B.
- Who it's for
- Developers and companies building autonomous AI agents that need to handle complex, professional tasks in a computer terminal.
- What it does
- It creates high-quality training data that teaches AI how to handle 'long-horizon' tasks, which are complex jobs requiring exploration, trial-and-error, and strategy adjustments. This allows the AI to move beyond simple commands to completing full professional projects autonomously.
- Interesting numbers
- Substantial performance gains on 12 benchmarks spanning coding and production settings
- Ability to continue experimenting and improving deliverables within a 12-hour research budget
- Explain it like I'm five
- Most AI agents are like students who only study the answer key; they know the right answer but not how to struggle through the problem. This research creates a 'practice gym' that forces the AI to explore, make mistakes, and fix them. It's like teaching a chef not just to follow a recipe, but how to taste the soup and adjust the seasoning until it's perfect.
#10 paper
DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models
Zhenglin Wan, Jiayu Chen, Zhengfeiyang Zhang +6
148130
#10 paper
DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models
Zhenglin Wan, Jiayu Chen, Zhengfeiyang Zhang +6
- What it is
- DiffusionART is a new training method used to fine-tune diffusion models (the AI that generates images and videos) using reinforcement learning.
- Who it's for
- AI researchers and developers who want to improve the quality of image and video generators without spending massive amounts of computing power.
- What it does
- It allows AI models to learn from 'black-box' rewards—scores given to an image without the AI knowing exactly why it got that score—more stably and efficiently. It removes the need for tedious manual tuning and reduces the amount of GPU power required to reach high-quality results.
- Interesting numbers
- Reached late-stage reward levels of DiffusionNFT using 12.02 times fewer GPU hours
- Reached late-stage reward levels of Flow-GRPO-Fast using 11.72 times fewer GPU hours
- Achieved 28.8% and 53.0% higher mean training rewards compared to DiffusionNFT and Flow-GRPO-Fast
- Explain it like I'm five
- Imagine trying to teach an artist to paint better, but you can only tell them 'this painting is a 7/10' without explaining which brushstroke was wrong. DiffusionART is like a smarter coach that figures out the general direction of improvement for the artist much faster, meaning the artist learns to paint masterpieces using way less practice time.
Trending models
on the Hugging Face hub#1 model · feature-extraction
google/embeddinggemma-2
apache-2.0 license
2.4k144.3k
#1 model · feature-extraction
google/embeddinggemma-2
apache-2.0 license
- What it is
- EmbeddingGemma 2 is an open-weights multimodal embedding model created by Google DeepMind that converts different types of data into mathematical vectors.
- Who it's for
- It is designed for developers building search or classification tools who want to run the model on consumer hardware, such as laptops and mobile devices.
- What it does
- The model maps text, images, video, and audio into a single shared mathematical space, allowing a computer to compare a piece of audio to a piece of text. It is notable for its 'modular' design, meaning developers can load only the specific encoders (like just audio or just vision) they need to save memory.
- Interesting numbers
- achieves a ~14% improvement on code tasks relative to its predecessor
- enabling up to a 6x reduction in vector storage costs with minimal impact on quality
- Explain it like I'm five
- Imagine a giant library where every book, photo, and song is assigned a specific GPS coordinate based on its meaning. If a song about 'sadness' and a poem about 'grief' have coordinates very close to each other, the computer knows they are related. This model is the tool that calculates those coordinates for almost any type of media.
#2 model · text-generation
jialinyyzz/humanizer
apache-2.0 license
1.7k75.3k
#2 model · text-generation
jialinyyzz/humanizer
apache-2.0 license
- What it is
- Developed by jialinyyzz, this is a text-generation model based on Google's Gemma-4-12B that has been fine-tuned (specifically trained) for text rewriting.
- Who it's for
- It is for people who want to make AI-generated text sound more natural in English or Chinese; it is designed to run locally on personal hardware, including laptops with as little as 8 GB of memory.
- What it does
- The model rewrites AI drafts—like emails or essays—to make them read like they were written by a human. Notably, it is trained to strictly preserve all dates, names, and numbers without adding new information, and it was created without using AI detectors during training.
- Explain it like I'm five
- Imagine you have a robot that writes perfectly correct but very stiff and boring essays. This model acts like a professional editor who rewrites those essays to sound natural and breezy, while making sure not a single fact or number is changed.
#3 model · text-to-image
Qwen/Qwen-Image-2.1-Turbo
other license
88015.3k
#3 model · text-to-image
Qwen/Qwen-Image-2.1-Turbo
other license
- What it is
- Qwen-Image-2.1-Turbo is an AI model created by the Qwen team that generates and edits images based on text descriptions.
- Who it's for
- It is designed for developers and AI researchers who have access to a CUDA-compatible GPU (a powerful graphics card) to run the model locally.
- What it does
- This model creates images and performs image editing using a 'turbo' process that requires only 8 denoising steps—the iterative process the AI uses to turn random noise into a clear picture. This makes it significantly faster than standard models while maintaining the architecture of the larger 7B (7 billion parameter) version.
- Explain it like I'm five
- Imagine an artist who usually takes 50 sketches to get a drawing perfect, but has now learned a shortcut to do it in just 8. This model is like that 'turbo' artist: it creates high-quality images from your words much faster than usual without needing to be told exactly how to sketch each step.
#4 model · image-text-to-text
openbmb/MiniCPM-V-4.7-35B-A3B
apache-2.0 license
7949.3k
#4 model · image-text-to-text
openbmb/MiniCPM-V-4.7-35B-A3B
apache-2.0 license
- What it is
- Created by OpenBMB, MiniCPM-V 4.7 is a multimodal AI model, meaning it can understand and discuss images, multiple photos, and videos.
- Who it's for
- It is designed for developers and users who want to run powerful AI locally on their own devices, specifically supporting iOS, Android, HarmonyOS, and various computer chips from Intel, AMD, and Arm.
- What it does
- It uses a 'Mixture of Experts' (MoE) architecture—a method where only a small part of the model activates for each task—to provide high performance without needing massive computing power. It is notable for its ability to run smoothly on mobile devices while maintaining a high level of visual understanding.
- Interesting numbers
- decoding speed of 13+ tokens/s with just 3 GB of memory on iPhone 17 Pro Max
- Explain it like I'm five
- Imagine a giant encyclopedia that is so smart it can 'see' and explain photos, but it's too big to fit in your pocket. This model is like a magic version of that book that only opens the specific pages it needs for each question, allowing it to fit and run fast on a smartphone.
#5 model · text-to-image
abenzerps/Qwen-Image-2.1-Uncensored-Quantized
other license
4.4k2.4M
#5 model · text-to-image
abenzerps/Qwen-Image-2.1-Uncensored-Quantized
other license
- What it is
- This is a modified version of the Qwen-Image-2.1 model, which is an AI designed to generate images from text descriptions.
- Who it's for
- It is for hobbyists who want to run image generation locally on their own hardware, including those with Apple Silicon (Macs) or NVIDIA GPUs.
- What it does
- The model creates images based on text prompts and is 'uncensored,' meaning it has fewer restrictions on what it will generate. It has been 'quantized,' a process that compresses the model so it takes up less memory and runs faster on home computers.
- Explain it like I'm five
- Imagine a giant, professional art studio that is too big to fit in your house. This model is like a 'miniaturized' version of that studio that keeps most of the talent but fits on your laptop. Plus, it's like an artist who doesn't have a boss telling them what they aren't allowed to paint.
#6 model · text-to-speech
fishaudio/drama-3
other license
4784.0k
#6 model · text-to-speech
fishaudio/drama-3
other license
- What it is
- Fish Audio Drama 3 is a text-to-speech model created by Fish Audio that turns written text into spoken audio.
- Who it's for
- This is for developers and content creators who want to build high-quality audio dramas or long-form spoken content using professional-grade AI hardware.
- What it does
- It can clone voices from short clips and handle complex scenes with up to 16 different characters in one text. It is notable for its 'emotional speech generation,' allowing users to insert tags for sounds like laughing or sighing, and its ability to edit specific words in a recording while keeping the rest of the audio consistent.
- Explain it like I'm five
- Imagine a digital voice actor that can play 16 different roles in one script and perfectly mimic anyone's voice from a tiny sample. Instead of just reading words, it can 'act' by adding sighs or laughs exactly where you tell it to. It's like a movie studio in a box that lets you change one word in a sentence without having to re-record the whole scene.
#7 model · image-to-video
Lightricks/LTX-2.5
other license
8.0k1.9M
#7 model · image-to-video
Lightricks/LTX-2.5
other license
- What it is
- LTX-2.5 is a generative AI model created by Lightricks that specializes in creating and editing video content.
- Who it's for
- This model is for digital creators and developers who want to generate high-quality video clips using their own hardware or tools like ComfyUI.
- What it does
- It is a versatile tool capable of turning text, images, or existing videos into new video clips. It is particularly notable for its wide range of inputs, including the ability to sync audio with video and support multiple languages.
- Explain it like I'm five
- Imagine a digital puppet master that can take a still photo or a written description and breathe life into it, turning it into a moving movie. It's like having a movie studio in a box that can change the scenery, add sound, or rewrite a scene just by following your instructions.
#8 model · text-to-image
AtomicChat/Qwen-Image-2.1-Turbo-Uncensored-GGUF
apache-2.0 license
490296.3k
#8 model · text-to-image
AtomicChat/Qwen-Image-2.1-Turbo-Uncensored-GGUF
apache-2.0 license
- What it is
- This is a modified version of the Qwen3-VL-8B-Instruct model, created by AtomicChat, which serves as a text encoder for generating images.
- Who it's for
- It is for users who want to generate images locally on their own hardware using tools like stable-diffusion.cpp.
- What it does
- The model uses a technique called 'abliteration' to remove the internal 'refusal direction,' meaning it won't lecture the user or refuse a prompt based on safety guidelines. While it removes these refusals in the text stage, it cannot force the image generator to draw things it was never trained to create.
- Interesting numbers
- English refusals dropped from 88.9% to 1.2%
- Russian refusals dropped from 38% to 0%
- MMLU (general knowledge) remained unchanged at 77.35%
- Explain it like I'm five
- Imagine a translator who usually refuses to translate 'rude' words. This model is like that translator after they've had a brain-switch that removes their 'refusal' instinct. Now, they will translate any prompt you give them into a language the image-maker understands, without arguing first.
#9 model · text-generation
LiquidAI/LFM2.5-24B-A2B
other license
3906.3k
#9 model · text-generation
LiquidAI/LFM2.5-24B-A2B
other license
- What it is
- LFM2.5-24B-A2B is a text-generation reasoning model created by Liquid AI.
- Who it's for
- It is designed for developers building private AI agents that run directly on local hardware, such as an Apple M5 Max or NVIDIA GPUs.
- What it does
- This model is built for 'agentic' tasks, meaning it is skilled at using tools to complete goals. It uses a Mixture of Experts (MoE) architecture—a design where only a small fraction of the model's total knowledge is used for any single task—to provide high-speed reasoning without needing massive computing power.
- Interesting numbers
- Scored 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B
- Runs at 184 tok/s on an Apple M5 Max
- Up to 21K output tok/s on a single NVIDIA B300 GPU
- Explain it like I'm five
- Imagine a giant library with 24 billion books, but instead of searching every shelf, the AI has a smart index that tells it exactly which 2 billion pages to read for your specific question. This makes it think deeply like a large model but respond as fast as a small one. It's like having a genius assistant that fits inside your laptop instead of a giant server farm.
#10 model · text-generation
ConwayResearch/Underdog-Saluki-27B-1.0
apache-2.0 license
497344.2k
#10 model · text-generation
ConwayResearch/Underdog-Saluki-27B-1.0
apache-2.0 license
- What it is
- Underdog Saluki 27B 1.0 is an open-weights text-generation model created by Underdog, based on the Qwen3.8-27B architecture.
- Who it's for
- It is designed for users who want a powerful model that can run on consumer-grade hardware, as it has been shrunk to under 8 GB to fit on standard laptops or home computers.
- What it does
- The model specializes in 'tool calling,' which is the ability for an AI to use external software functions to complete tasks. It is notable for maintaining high performance in these tasks despite being significantly smaller than the original full-sized model.
- Interesting numbers
- Passed 88 of 120 tool calling tasks compared to 84 for the full-size model
- Reduced size from 54 GB down to 7.89 GB
- Achieved 42 parallel tool calls versus 35 for the full-size model
- Explain it like I'm five
- Imagine taking a massive, 50-pound encyclopedia and shrinking it down to a pocket-sized notebook without losing the important facts. Underdog Saluki does this for AI, making a huge model small enough to fit on a normal computer while keeping its ability to 'use tools' (like a calculator or a calendar) intact.