Research tracking
Trending papers and models, explained simply
Trending papers
this week#1 paper
LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering
Yi Wang, Haopeng Zhang, Chengxiang Huang +4
8267
#1 paper
LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering
Yi Wang, Haopeng Zhang, Chengxiang Huang +4
- What it is
- LoopArena is a benchmark, which is a standardized test used to measure and compare the performance of AI systems.
- Who it's for
- AI researchers and developers building autonomous coding agents that need to manage complex, multi-step software projects.
- What it does
- It separates the 'brain' that manages the project (the Controller) from the 'hands' that write the code (the Worker). This allows researchers to test if a model is actually good at planning and monitoring progress, rather than just being good at writing a single snippet of code.
- Interesting numbers
- Best observed Strict Success Rate on full tasks was 24.69%
- Average reduction in estimated inference cost across Controllers was 64.4%
- Explain it like I'm five
- Imagine a construction site where one person is the Manager and another is the Builder. Usually, we only check if the house was finished, but we don't know if the Manager gave bad orders or if the Builder just messed up. LoopArena is like a test that specifically grades the Manager on how well they organize the work and spot mistakes.
#2 paper
DART-SD: Diamond-topology Aware Retrieval and Tuning for Self-Distillation of Multi-Turn Tool-Calling Agents
Hangrui Xu, Jiarui Wang, Yang Yang +5
60
#2 paper
DART-SD: Diamond-topology Aware Retrieval and Tuning for Self-Distillation of Multi-Turn Tool-Calling Agents
Hangrui Xu, Jiarui Wang, Yang Yang +5
- What it is
- DART-SD is a new training framework designed to improve how AI agents use tools over multiple steps to solve complex tasks.
- Who it's for
- Developers and researchers building autonomous AI agents that need to interact with external software or APIs to complete goals.
- What it does
- It stops the AI from being forced to follow one single 'correct' path and instead allows it to explore multiple valid ways to reach a goal. When the AI makes a mistake, the system identifies the exact point of failure and teaches it how to recover using successful examples.
- Explain it like I'm five
- Imagine learning to cook a meal. Most AI training forces you to follow one exact recipe, and if you swap the order of two steps, the AI thinks you failed. DART-SD is like a coach who realizes there are many ways to cook the meal; if you mess up, the coach points to the exact moment you went wrong and shows you how to get back on track.
#3 paper
Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models
Senqiao Yang, Chengyao Wang, Yuxin Chen +13
593.6k
#3 paper
Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models
Senqiao Yang, Chengyao Wang, Yuxin Chen +13
- What it is
- A new training method and model backbone called VLAct designed for Vision-Language-Action (VLA) models, which are AI systems that translate visual and text inputs into physical robot movements.
- Who it's for
- Robotics researchers and companies building general-purpose robots that need to learn new tasks quickly without needing massive amounts of expensive, hand-collected data.
- What it does
- It improves how robots understand the relationship between what they see and how they move by pre-training on diverse data from many different types of robots. This allows the AI to develop a 'general sense' of action that transfers easily to new robots or tasks, reducing the amount of specific training data required.
- Interesting numbers
- Achieved success rates of 82.6% on LIBERO-Plus and 92.5% on RoboTwin 2.0
- On RoboCasa-GR1, outperformed the GR00T-N1.6 baseline while using only 20% of the downstream trajectories
- Explain it like I'm five
- Imagine trying to teach a robot to cook, but you only have a few videos of it happening. Instead of just memorizing those few videos, VLAct is like giving the robot a 'general education' in how all kinds of machines move first. Because it already understands the basics of motion, it can learn a specific new task—like flipping a pancake—much faster and with far fewer examples.
#4 paper
Agentic Artifact Creation: Systems, Evaluation, Principles, and Opportunities
Tianfu Wang, Zhezheng Hao, Xilin Xia +9
5114
#4 paper
Agentic Artifact Creation: Systems, Evaluation, Principles, and Opportunities
Tianfu Wang, Zhezheng Hao, Xilin Xia +9
- What it is
- This is a survey paper that reviews and organizes existing research on how AI agents build complex, finished projects.
- Who it's for
- AI researchers and developers who are building systems that need to create reliable, professional-grade deliverables rather than just simple drafts.
- What it does
- It defines a framework for 'agentic artifact creation,' where an AI doesn't just generate a response but iteratively builds, tests, and fixes a project based on feedback. The paper analyzes 259 different works to establish principles for making AI-created work more dependable and accountable.
- Explain it like I'm five
- Imagine the difference between an AI that just writes a recipe and an AI that actually cooks the meal, tastes it, realizes it's too salty, and adds a potato to fix it. This paper studies how to move AI from just 'suggesting' things to actually 'building and fixing' complete projects until they are perfect.
#5 paper
Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning
Hanyang Wang, Yimo Cai, Weiliang Chen +14
43263
#5 paper
Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning
Hanyang Wang, Yimo Cai, Weiliang Chen +14
- What it is
- A new framework called Code-as-World that represents physical environments as executable computer code rather than just images or text.
- Who it's for
- AI researchers and developers building models that need to understand physics, robotics, and how the physical world evolves over time.
- What it does
- It uses an AI agent to turn videos or descriptions into a runnable simulation by proposing, testing, and refining code until it matches reality. These verified simulations are then used to train vision-language models to better reason about quantitative physical properties.
- Interesting numbers
- achieves state-of-the-art performance on QuantiPhy
- surpasses leading proprietary models
- Explain it like I'm five
- Imagine trying to teach a robot how a ball bounces. Instead of just showing it a thousand videos, this system writes a mini-physics program (like a simple video game) that perfectly mimics that bounce. By playing with this 'code version' of the world, the AI learns the actual rules of physics instead of just guessing based on pictures.
#6 paper
J-Zero: Unified Challenger--Solver--Judge Co-Evolution from Zero Data
Gyouk Chu, Myeongho Jeon, Eunho Yang
352
#6 paper
J-Zero: Unified Challenger--Solver--Judge Co-Evolution from Zero Data
Gyouk Chu, Myeongho Jeon, Eunho Yang
- What it is
- J-Zero is a new training framework that allows AI models to improve themselves without needing new data from humans.
- Who it's for
- AI researchers and companies building large language models that need to get smarter without expensive human labeling.
- What it does
- It creates a loop where one AI (the Challenger) makes harder problems, another (the Solver) tries to solve them, and a third (the Judge) learns to grade them. This process allows the AI to improve in both 'verifiable' tasks (like math, where there is one right answer) and 'unverifiable' tasks (like writing, where quality is subjective).
- Interesting numbers
- Outperforms baselines by an average of 4.2 points on verifiable domains
- Outperforms baselines by an average of 8.0 points on unverifiable domains
- Continues to improve through at least ten iterations, whereas baselines degrade after two
- Explain it like I'm five
- Imagine a student, a teacher, and a test-writer who are all AI. The test-writer keeps making harder exams, the student studies to pass them, and the teacher learns how to grade them better. Because they all push each other to be better, the AI gets smarter on its own without needing a human to give it the answers.
#7 paper
Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction
Jiarong Han, Jincheng Xiong, Yuzhou Liu +6
28298
#7 paper
Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction
Jiarong Han, Jincheng Xiong, Yuzhou Liu +6
- What it is
- A new streaming AI model called ABot-Recon designed for 3D reconstruction from very long videos.
- Who it's for
- Researchers and developers building systems for autonomous navigation, robotics, or augmented reality that need to map environments in real-time.
- What it does
- It creates 3D maps and tracks camera movement using only a small window of recent frames to prevent the system from slowing down or crashing as videos get longer. It specifically reduces 'drift,' which is the gradual accumulation of small errors that make a 3D map become distorted over time.
- Interesting numbers
- On Oxford Spires, achieves an ATE (Absolute Trajectory Error) of 4.35 m
- Achieves an RPE-R (Relative Pose Error) of 0.12°
- Reduces both errors by approximately 40% relative to the best prior results
- Explain it like I'm five
- Imagine trying to draw a map of a city while walking through it. Instead of trying to remember every single street you've passed since you started (which would eventually overwhelm your brain), this AI only looks at the last few steps you took to figure out where you are. By focusing on these small, local chunks and using a special 'refiner' to double-check its work, it keeps the map accurate without getting confused by the length of the journey.
#8 paper
Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090
Kairong Luo, Jiarui Cui, Yaorui Yin +8
2648
#8 paper
Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090
Kairong Luo, Jiarui Cui, Yaorui Yin +8
- What it is
- A cost-efficient training recipe and a collection of small language models called Puro-2B.
- Who it's for
- Academic researchers and open-source developers who lack the massive budgets typically required to train AI from scratch.
- What it does
- It provides a blueprint to train high-performing AI models using consumer-grade hardware (like the RTX 5090 GPU) rather than expensive industrial server farms. By using low-precision math and optimized data strategies, it drastically lowers the financial barrier to creating custom AI models.
- Interesting numbers
- Best model trained for less than $6.9K
- Performance of Qwen2-1.5B can be reached with less than $5,090
- Trained on up to 1.4 trillion tokens
- Explain it like I'm five
- Usually, building a powerful AI is like building a skyscraper—it costs millions of dollars and requires massive machinery. This research is like finding a way to build a high-quality tiny house using tools you can buy at a local hardware store for a few thousand dollars. It proves that you don't need a billion-dollar company to create a smart AI model.
#9 paper
LayerRecall: A State-Conditioned Memory Router for Long-Horizon Consistency in Video Generation
Yixuan Ding, Jiahao Kong, Wei Huang +2
256
#9 paper
LayerRecall: A State-Conditioned Memory Router for Long-Horizon Consistency in Video Generation
Yixuan Ding, Jiahao Kong, Wei Huang +2
- What it is
- LayerRecall is a new memory routing framework designed to improve how AI models remember details over long periods in generated videos.
- Who it's for
- Researchers and developers building AI video generators that need to maintain consistent characters and scenes across long clips.
- What it does
- It allows a video model to selectively retrieve and use specific historical data only in the layers of the model that need it most. This prevents the AI from forgetting what a character or object looked like at the start of a video while still keeping the immediate motion smooth.
- Interesting numbers
- achieves the best overall results on MemoBench and MovieBench
- matching its backbone on VBench-Long
- Explain it like I'm five
- Imagine an AI drawing a movie frame-by-frame. Usually, it only remembers the last few seconds, so a character's shirt might accidentally change color after a minute. LayerRecall acts like a smart filing cabinet that tells the AI exactly which old memories to look at and where to apply them, ensuring the character stays the same from the first scene to the last.
#10 paper
ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL
Zhuoshi Pan, Qizhi Pei, Junru Lu +4
2416
#10 paper
ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL
Zhuoshi Pan, Qizhi Pei, Junru Lu +4
- What it is
- ContextPilot is a framework that teaches AI agents how to proactively manage their own working memory during long, complex tasks.
- Who it's for
- Developers and companies building AI agents that need to handle massive amounts of information over long conversations without getting confused or slowing down.
- What it does
- It gives AI models a better set of tools to plan, remember long-term data, and compress information. It also uses a specialized training method called Reinforcement Learning (RL)—a way of learning through trial and error—to help the AI decide exactly when and how to edit its memory for the best result.
- Interesting numbers
- Outperforms existing baselines across various base models and benchmarks in long-context QA and deep search tasks
- Explain it like I'm five
- Imagine an AI is like a student taking a massive open-book exam. Instead of letting their desk get buried under a mountain of papers, ContextPilot teaches the student how to take organized notes, throw away useless scraps, and keep only the most important clues in sight. This keeps the AI from getting overwhelmed and helps it find the right answer faster.
Trending models
on the Hugging Face hub#1 model · image-text-to-text
Qwen/Qwen3.8-Flash-Next
other license
4.5k158.6k
#1 model · image-text-to-text
Qwen/Qwen3.8-Flash-Next
other license
- What it is
- Qwen3.8-Flash-Next is an experimental AI model from the Qwen team that can process both images and text to generate text responses.
- Who it's for
- This model is for AI researchers and developers who want to test the new architecture that will power the future Qwen4 models.
- What it does
- It is designed for high efficiency, using a 'sparse attention' system to handle long conversations more quickly. It also uses a unique 'n-gram embedding' method to increase the model's knowledge without requiring massive amounts of extra computing power.
- Explain it like I'm five
- Imagine a library where, instead of reading every single page to find an answer, the librarian has a super-fast shortcut to the exact paragraph they need. This model uses similar shortcuts to process huge amounts of information much faster than older AI, making it 'flashy' and efficient.
#2 model · image-text-to-text
zai-org/GLM-5.3-Flash
mit license
1.8k379.3k
#2 model · image-text-to-text
zai-org/GLM-5.3-Flash
mit license
- What it is
- GLM-5.3-Flash is an open-weights AI model created by Z.ai that is natively multimodal, meaning it can process both text and images.
- Who it's for
- It is designed for developers and researchers who want a high-performance model they can run locally using specialized frameworks like vLLM or SGLang.
- What it does
- The model is built for efficiency, using a 'sparse' architecture that only activates a small fraction of its total parameters (18B out of 320B) to save on computing costs. It excels at coding and 'agentic' tasks—tasks where the AI must use tools to solve complex problems—while remaining cheaper to run than previous versions.
- Interesting numbers
- Outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price
- Approaching Claude Opus 4.8 on coding and agentic benchmarks
- Explain it like I'm five
- Imagine a giant library with 320 billion books, but the librarian is so smart they only need to open 18 billion of them to answer your question. This makes the AI much faster and cheaper to run without losing its intelligence. It's like having a super-genius assistant that can 'see' images and write code, but doesn't need a massive power plant to operate.
#3 model · text-generation
zai-org/GLM-5.3
other license
1.4k66.2k
#3 model · text-generation
zai-org/GLM-5.3
other license
- What it is
- GLM-5.3 is an open-weights text-generation model created by Z.ai that can communicate in both English and Chinese.
- Who it's for
- This model is for developers and security researchers who want to run a powerful coding assistant on their own servers using frameworks like vLLM or SGLang.
- What it does
- It specializes in complex coding and 'cyber capabilities,' such as finding and exploiting software vulnerabilities. Notably, it achieves these gains through post-training—refining the model's behavior after its initial learning phase—rather than changing the base architecture.
- Interesting numbers
- Achieved a score of 84.5 on CyberGym for vulnerability discovery
- More than doubled the performance of GLM-5.2 on exploitation benchmarks
- Scored 48.2 on AutomationBench (v1.0.6)
- Explain it like I'm five
- Imagine a student who already knew the basics of computer science but then spent months in an intensive 'bootcamp' for hacking and advanced coding. GLM-5.3 is like that student; it uses the same brain as the previous version but has become a specialist in solving the hardest technical puzzles.
#4 model · image-text-to-text
Qwen/Qwen3.8-27B
apache-2.0 license
13.5k4.7M
#4 model · image-text-to-text
Qwen/Qwen3.8-27B
apache-2.0 license
- What it is
- Qwen3.8-27B is an open-weights AI model from the Qwen team that can process both text and visual information, such as images and videos.
- Who it's for
- It is designed for developers and researchers who want a powerful but 'compact' model that can be deployed on their own infrastructure using tools like vLLM or SGLang.
- What it does
- The model excels at coding, professional research, and 'agentic tasks,' which are complex jobs where the AI must plan and act autonomously. It is notable for its 'flexible thinking control,' allowing users to tune how deeply the model reasons through a problem.
- Explain it like I'm five
- Imagine a digital assistant that doesn't just read your text, but can also watch a hour-long video or study a science diagram to help you. It's like a student who can switch between 'quick answer mode' and 'deep thinking mode' depending on how hard the homework is.
#5 model · image-text-to-text
unsloth/Qwen3.8-Flash-Next-GGUF
other license
631373.0k
#5 model · image-text-to-text
unsloth/Qwen3.8-Flash-Next-GGUF
other license
- What it is
- Qwen3.8-Flash-Next is an experimental open-weights model from the Qwen team that can process both images and text to generate text responses.
- Who it's for
- It is designed for developers and AI enthusiasts who want to run a high-performance model on their own hardware, specifically using tools like Unsloth Desktop or llama.cpp.
- What it does
- This model serves as a preview for the upcoming Qwen4 architecture, focusing on extreme efficiency. It uses new methods to handle long conversations faster and manages memory more effectively than traditional models.
- Explain it like I'm five
- Imagine a giant library where the librarian usually has to read every single page to find an answer. This model uses a new 'shortcut' system to skip the fluff and find the right information instantly, making it much faster and lighter without losing its intelligence.
#6 model
unsloth/Qwen3.8-27B-GGUF
apache-2.0 license
3.3k9.1M
#6 model
unsloth/Qwen3.8-27B-GGUF
apache-2.0 license
- What it is
- This is Qwen3.8-27B, an open-weights AI model from the Qwen family that can process both text and visual data like images and videos.
- Who it's for
- It is designed for developers and researchers who want a powerful model that can run on their own hardware, specifically those using Unsloth tools to run it on a desktop.
- What it does
- The model excels at coding, professional research, and 'agentic tasks,' which are complex jobs where the AI must plan and act autonomously. It also features a 'thinking mode' that allows users to control how deeply the AI reasons through a problem before answering.
- Explain it like I'm five
- Imagine a digital brain that doesn't just read text, but can also 'watch' a movie or 'look' at a diagram to understand it. It's like having a very smart assistant that can stop and think carefully before speaking, making it great for solving hard puzzles or writing computer code.
#7 model · text-generation
tencent/Hy4-preview
apache-2.0 license
3492.6k
#7 model · text-generation
tencent/Hy4-preview
apache-2.0 license
- What it is
- Hy4-preview is a large-scale text-generation AI model created by the Tencent Hy Team.
- Who it's for
- This model is for advanced developers and researchers with access to high-end industrial hardware, as it is a massive 'flagship' model with 770 billion total parameters.
- What it does
- It is a Mixture-of-Experts (MoE) model, meaning it contains many specialized sub-networks but only activates a small fraction of them for each task to save computing power. It is notable for its massive 1-million-token context length, allowing it to process enormous amounts of text at once.
- Explain it like I'm five
- Imagine a library with 770 billion books, but instead of reading every book to answer a question, the AI has a smart manager who only picks the 49 billion most relevant pages. This makes it incredibly knowledgeable but efficient enough to actually run. It also has a 'super-memory' that lets it remember a massive amount of information from a very long conversation.
#8 model · image-to-video
Lightricks/LTX-2.5
other license
2.4k1.2M
#8 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.
#9 model · image-text-to-text
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
mit license
3240
#9 model · image-text-to-text
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
mit license
- What it is
- DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal AI model created by DeepSeek that can process both text and images.
- Who it's for
- This model is for developers and AI researchers who want to build 'agents'—AI programs that can take actions—and who have the technical setup to run open-weights models via the transformers library.
- What it does
- It combines a fast text-processing architecture with new visual modules to understand images and charts. It is specifically designed to improve 'multimodal agent' capabilities, meaning it can better use visual information to complete complex tasks without losing its ability to handle text-only jobs.
- Interesting numbers
- ApexBench (Pass@1): 36.5 (compared to 26.2 for the text-only version)
- Chartography: 64.3
- ZeroBench (Pass@5): 35.0
- Explain it like I'm five
- Imagine a smart assistant that could previously only read instructions but now has eyes. This model is like giving that assistant a pair of glasses so it can look at a chart or a screen and actually use that information to help you get work done.
#10 model · text-generation
unsloth/GLM-5.3-Flash-GGUF
mit license
30953.4k
#10 model · text-generation
unsloth/GLM-5.3-Flash-GGUF
mit license
- What it is
- GLM-5.3-Flash is a multimodal AI model created by Z.ai that can process and generate text in both English and Chinese.
- Who it's for
- It is for developers and AI enthusiasts who want to run a powerful model on their own hardware, including those using the Unsloth Desktop app to run it on a personal computer.
- What it does
- The model uses a 'sparse' architecture, meaning it only activates a small fraction of its total brain power (18 billion out of 320 billion parameters) at one time to save energy and speed up responses. It is designed to be highly efficient at coding and acting as an 'agent' that can complete complex tasks.
- Interesting numbers
- Approaches Claude Opus 4.8 on coding and agentic benchmarks
- Outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price
- Explain it like I'm five
- Imagine a giant library with 320 billion books, but instead of searching every single shelf for an answer, the AI has a shortcut that lets it only look at the 18 billion most relevant pages. This makes it incredibly fast and cheap to run without losing its intelligence. It's like having a super-genius assistant that doesn't need a massive supercomputer to think.