Research tracking
Trending papers and models, explained simply
Trending papers
this week#1 paper
BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
Björn Engdahl, Adrian Kosowski, Jan Chorowski +6
5531.1k
#1 paper
BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
Björn Engdahl, Adrian Kosowski, Jan Chorowski +6
- What it is
- A new AI reasoning model called BDH-CQ that combines the ability to learn from examples on the fly with a hidden, internal loop for thinking.
- Who it's for
- AI researchers and developers working on general intelligence and efficient problem-solving models.
- What it does
- It allows a model to solve complex puzzles by processing information into a 'latent space'—a mathematical internal representation—where it can iterate on a solution without having to write out its thoughts in words. This makes the reasoning process significantly cheaper and faster while maintaining accuracy.
- Interesting numbers
- 29.5% pass@2 on the ARC-AGI-1 evaluation set
- Inference cost of $0.0007 per task
- Explain it like I'm five
- Imagine if, instead of solving a math problem by writing out every single step on paper, you could just 'think it through' in your head until you found the answer. This AI does exactly that; it uses a hidden internal loop to reason through puzzles, making it much faster and cheaper than models that have to talk through their logic.
#2 paper
Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA
Mind Lab, Vin Bo, Asher Cai +73
324
#2 paper
Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA
Mind Lab, Vin Bo, Asher Cai +73
- What it is
- Macaron-V1 is a family of AI agent models designed for 'continual learning,' which means the AI can keep learning from new experiences even after it has been released.
- Who it's for
- Developers and companies building advanced AI agents that need to handle specialized tasks like coding, chatting, and creating user interfaces.
- What it does
- It uses a 'Mixture-of-LoRA' architecture, which keeps a main brain frozen while adding small, specialized plug-ins (called LoRAs) for different tasks. This allows the system to improve itself recursively by evaluating its own performance and updating its versions over time.
- Explain it like I'm five
- Imagine a student who has a giant textbook for general knowledge but also keeps a collection of small, specialized notebooks for math, art, and coding. Instead of trying to rewrite the whole textbook every time they learn something new, they just update the specific notebook that needs it. Macaron-V1 does this with AI, allowing it to get smarter at specific skills without forgetting everything else.
#3 paper
On-Policy Self-Distillation without Any Supervision
Yijiang Li, Bingyang Wang, Yijun Liang +3
185
#3 paper
On-Policy Self-Distillation without Any Supervision
Yijiang Li, Bingyang Wang, Yijun Liang +3
- What it is
- A new training method called U-OPSD that allows Large Language Models to improve their own performance without needing external answers or human feedback.
- Who it's for
- AI researchers and developers who want to make models smarter without the high cost of collecting labeled data or using larger, expensive 'teacher' models.
- What it does
- It enables a model to find its own mistakes by comparing multiple attempts at the same problem. By identifying where it is consistently wrong and correcting those specific errors, the model effectively teaches itself to be more accurate.
- Interesting numbers
- Improves over the base model by 8.5% and 10.7% on Qwen3 non-thinking mode at 4B and 8B scales
- Outperforms OPSD by 3.2% and 2.3% on average across five math benchmarks
- Surpasses GRPO by 0.7% and 1.1% in thinking mode
- Explain it like I'm five
- Imagine a student practicing math problems without an answer key. They solve the same problem ten times; if they get the same answer eight times, they assume that's the right one and use it to fix the two times they got it wrong. By doing this over and over, the student gets smarter just by checking their own consistency.
#4 paper
ComBodied Agents: a New Paradigm of Human-Centric Agentic AI
Qianggang Ding, Xingyao Wang, Rui Feng +19
173
#4 paper
ComBodied Agents: a New Paradigm of Human-Centric Agentic AI
Qianggang Ding, Xingyao Wang, Rui Feng +19
- What it is
- A new conceptual framework called 'ComBodied Agents' that proposes a human-centric approach to AI design.
- Who it's for
- Researchers and developers building personal assistants, health-care AI, and robotic companions.
- What it does
- It moves AI beyond just completing tasks to instead focusing on the human's well-being over time. It integrates digital tools and physical robots into one system that can perceive a person's state, predict their needs, and provide the right kind of support.
- Explain it like I'm five
- Most AI is like a tool: one helps you with your screen (digital) and another helps you move things (robotic). This research proposes a 'brain' that connects both, acting more like a thoughtful caregiver who doesn't just remind you to take medicine, but understands why you forgot and helps you solve the actual problem.
#5 paper
SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring
Yuling Shi, Jinghan Xu, Kelin Fu +12
125
#5 paper
SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring
Yuling Shi, Jinghan Xu, Kelin Fu +12
- What it is
- A new benchmark, which is a standardized test used to measure the performance of AI models, specifically designed for large-scale code refactoring.
- Who it's for
- AI researchers and companies building autonomous coding agents that need a more rigorous way to test their software engineering skills.
- What it does
- It provides a curated set of 170 real-world tasks across seven programming languages that require changing code across multiple files without breaking existing functionality. It improves on previous tests by removing ambiguous instructions and flawed tests that previously gave AI models unfair advantages or incorrect failures.
- Interesting numbers
- Average of 11.4 modified files per instance
- Average of 261.6 lines of code per instance
- Best model achieves only 41.2% resolve rate
- Explain it like I'm five
- Imagine giving an AI a giant LEGO castle and asking it to change the foundation without knocking over the towers. Most current tests are too easy, like asking the AI to swap one brick; this new test forces the AI to reorganize huge sections of the build across many different rooms. Even the smartest AI models currently fail more than half the time.
#6 paper
Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design
Qing Zong, Jiayu Liu, Junhao Shen +9
11627
#6 paper
Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design
Qing Zong, Jiayu Liu, Junhao Shen +9
- What it is
- This is a survey paper, which means it is a comprehensive review and organization of existing research rather than a new model or tool.
- Who it's for
- AI researchers and developers building 'agentic systems'—AI programs designed to act independently to achieve goals.
- What it does
- It creates a roadmap for how AI agents can improve themselves without human help by interacting with other agents and changing environments. It categorizes this growth into three stages: agents learning from each other, agents adapting to their surroundings, and agents improving the very way they evolve.
- Explain it like I'm five
- Imagine a video game where the characters and the levels both learn and change based on how the other behaves. Instead of a human programmer writing every rule, the AI 'evolves' by challenging itself and its peers. This paper explains how we can move from AI that follows a fixed script to AI that grows and learns on its own.
#7 paper
Beyond Pixels: From Video Priors to 4D Worlds
Zihao Liu, Xiaolong Shen, Zhenglin Zhou +2
108
#7 paper
Beyond Pixels: From Video Priors to 4D Worlds
Zihao Liu, Xiaolong Shen, Zhenglin Zhou +2
- What it is
- A new framework called Latent-to-4D that generates dynamic 3D scenes (4D) directly from the internal data of video AI models.
- Who it's for
- Researchers and developers building 3D animation tools, virtual reality environments, or digital movie effects.
- What it does
- It removes the need to first create a 2D video and then try to turn it into 3D, which often causes errors. Instead, it translates the 'hidden' mathematical representations (latents) of a video model directly into 3D geometry, making the process more stable and compatible with different video generators.
- Interesting numbers
- Surpasses matched same-latent Wan+4RC cascades in projection-based DINO-F1 by 2.88--3.45 points on Text4D-200
- Surpasses matched same-latent Wan+4RC cascades in projection-based DINO-F1 by 5.81 points on I4D-200
- Explain it like I'm five
- Imagine trying to build a LEGO sculpture by looking at a photo of one; you might get the shapes wrong. This AI instead looks at the 'blueprint' inside the computer's mind before the photo is even made. This allows it to build a moving 3D world that is much more accurate and smooth.
#8 paper
SimWAM: A Simple World Action Model for End-to-End Autonomous Driving
Zongchuang Zhao, Xin Zhou, Tianyang Xu +5
10245
#8 paper
SimWAM: A Simple World Action Model for End-to-End Autonomous Driving
Zongchuang Zhao, Xin Zhou, Tianyang Xu +5
- What it is
- SimWAM is a new model architecture for end-to-end autonomous driving that uses video generation to help a car learn how to drive.
- Who it's for
- Researchers and companies building self-driving software who want more accurate trajectory planning without slowing down the car's reaction time.
- What it does
- It uses a 'video expert' during training to teach the AI how the world moves, but throws that expert away during actual driving to keep the system fast. This allows the car to predict its path directly without needing to generate expensive fake videos in real-time.
- Interesting numbers
- Achieves 91.5 PDMS on NAVSIM
- Surpasses state-of-the-art WAM-based planners with substantially lower latency
- Explain it like I'm five
- Imagine learning to ride a bike by watching thousands of videos of other people riding. SimWAM is like a student who studies those videos to understand how physics works, but when it's time to actually ride, they don't need to keep watching videos—they just use the 'gut feeling' they developed during study. This makes the AI smarter at driving but much faster at making decisions.
#9 paper
Stealing Reasoning Traces from Proprietary LLM APIs
Alexander Panfilov, David Schmotz, Ilia Shumailov +5
86
#9 paper
Stealing Reasoning Traces from Proprietary LLM APIs
Alexander Panfilov, David Schmotz, Ilia Shumailov +5
- What it is
- A security research paper that identifies a vulnerability in how major AI companies hide the internal 'reasoning' steps of their models.
- Who it's for
- AI providers like OpenAI, Google, and Anthropic, as well as security researchers and developers who handle AI session logs.
- What it does
- It demonstrates a method to steal hidden reasoning traces by moving encrypted data from a powerful model to a weaker one that is easier to trick into revealing the text. This allows attackers to bypass intellectual property protections, extract private user data, and uncover hazardous information the AI tried to hide.
- Interesting numbers
- decoded 315,320 reasoning blocks from public repositories
- recovered 367 Personally Identifiable Information (PII) artifacts
- recovered 182 credentials
- Explain it like I'm five
- Imagine a genius AI writes its secret thoughts in a coded diary and gives it to you, but you don't have the key. The researchers found that if you give that coded diary to a 'dumber' AI from the same company, that AI will accidentally read the secrets out loud for you. This lets people steal the genius AI's secrets or find private passwords hidden in the code.
#10 paper
Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution
Anton Razzhigaev, Andrei Gritsaev, Andrei Kaznacheev +3
791.1k
#10 paper
Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution
Anton Razzhigaev, Andrei Gritsaev, Andrei Kaznacheev +3
- What it is
- Ouroboros is a self-developing framework for AI coding agents that allows the AI to rewrite its own internal code, prompts, and tools to improve itself over time.
- Who it's for
- Software developers and AI researchers building autonomous agents that need to adapt and improve without constant manual updates from humans.
- What it does
- It enables an AI to identify its own bugs and inefficiencies and then propose code changes to its own core system. These changes are reviewed and then become the new permanent version of the AI, creating a cycle of continuous self-evolution.
- Interesting numbers
- Terminal-Bench 2.1: 86.74% (best result reported)
- OSWorld-Verified: 90.69% (exceeding best previously reported score)
- CL-Bench: normalized reward of 0.2301 (new state of the art)
- Explain it like I'm five
- Imagine a robot that comes with a manual on how to work, but the robot can actually rewrite that manual to make itself smarter. If it finds a better way to solve a problem, it updates its own brain so it never makes that mistake again. It's like a video game character that can rewrite its own code to unlock new abilities while it's playing.
Trending models
on the Hugging Face hub#1 model · image-text-to-text
meta-models/Muse-Glimmer-30B
apache-2.0 license
1.3k0
#1 model · image-text-to-text
meta-models/Muse-Glimmer-30B
apache-2.0 license
- What it is
- Muse Glimmer is an open-weights AI model created by Meta Superintelligence Lab that can process both text and images to generate text responses.
- Who it's for
- It is designed for developers and users who want to run a powerful AI agent locally on consumer-grade hardware, such as a high-end home computer, without needing an internet connection.
- What it does
- The model acts as an autonomous agent, meaning it can plan multi-step tasks, use digital tools, and fix its own mistakes if a tool fails. It is notable for its ability to 'see' screenshots and documents while reasoning through complex workflows in over 100 languages.
- Interesting numbers
- Strong success rates on DeepSearch QA, MCP-Atlas, 𝛕3-Bench and SWE-Bench
- Explain it like I'm five
- Imagine a digital assistant that doesn't just chat, but can actually 'do' things for you on your computer, like a tiny employee living in your hard drive. It can look at a picture of a chart, figure out a plan to analyze it, and if it hits a snag, it tries a different way instead of just giving up.
#2 model · image-text-to-video
MiniMaxAI/MiniMax-H3
other license
3.7k83.5k
#2 model · image-text-to-video
MiniMaxAI/MiniMax-H3
other license
- What it is
- MiniMax H3 is an 'omni-modal' generative system created by MiniMaxAI, meaning it can understand and create content across text, images, video, and audio simultaneously.
- Who it's for
- This is for developers and creative technologists who want to build high-end video generation tools that include synchronized sound.
- What it does
- It generates high-resolution videos (up to 2K) between 4 and 15 seconds long with built-in stereo audio. It is notable for its flexibility, allowing users to start a video from a text prompt, a single image, or even a pair of images representing the first and last frames.
- Explain it like I'm five
- Imagine a movie director who can take a few photos, a voice recording, and a written script, and instantly turn them into a short, high-quality movie clip with matching sound. Instead of just making a silent GIF, this AI handles the picture and the noise at the same time so they fit perfectly together.
#3 model · text-to-video
larryvrh/MiniMax-H3-Turbo-Lora
apache-2.0 license
7010
#3 model · text-to-video
larryvrh/MiniMax-H3-Turbo-Lora
apache-2.0 license
- What it is
- Created by user larryvrh, this is a LoRA (a small, lightweight add-on file) for the MiniMax-H3 model that generates synchronized video and stereo audio.
- Who it's for
- It is for AI artists and developers using ComfyUI (a visual tool for building AI workflows) who want to generate high-quality video and sound much faster.
- What it does
- It acts as a 'turbo' boost that allows the model to create video and audio in just 4 sampling steps instead of the usual 20. This provides roughly a 5x increase in speed, though it is currently a preview version that can sometimes produce 'plastic-looking' skin or grainy noise.
- Interesting numbers
- roughly a 5× speedup in sampling wall-clock
- renders joint video + synchronized stereo audio in 4 sampling steps instead of the usual ~20
- Explain it like I'm five
- Imagine if a digital artist usually took 20 brushstrokes to finish a painting, but you gave them a magic tool that let them do it in only 4 strokes. It's way faster, but because it's still a prototype, some of the details might look a bit too shiny or 'plastic' until the creator fixes it.
#4 model · text-generation
deepseek-ai/DeepSeek-V4-Flash-0731
mit license
3.2k1.0M
#4 model · text-generation
deepseek-ai/DeepSeek-V4-Flash-0731
mit license
- What it is
- DeepSeek-V4-Flash-0731 is a text-generation AI model created by DeepSeek AI.
- Who it's for
- This model is for developers and researchers who want a high-performance AI that is efficient to run, specifically those using 8-bit or fp8 precision hardware to save memory.
- What it does
- It is designed for 'agentic' tasks, meaning it is very good at using tools and solving complex coding problems autonomously. It uses a 'speculative decoding module,' which is a technique to speed up how quickly the AI generates text, allowing it to outperform larger models while using fewer active parameters.
- Interesting numbers
- Terminal Bench 2.1: 82.7
- DeepSWE: 54.4
- Toolathlon-Verified: 70.3
- DSBench-FullStack: 68.7
- Explain it like I'm five
- Imagine a giant encyclopedia that can think and write. Most 'smart' models are like heavy books that take a long time to flip through, but this 'Flash' version is like a high-speed cheat sheet that gives almost the same quality of answers much faster. It's especially good at acting like a digital assistant that can actually write and fix computer code on its own.
#5 model · image-to-video
Lightricks/LTX-2.5
other license
56839
#5 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.
#6 model · text-generation
Qwen/Qwen3.8-2.4T-A95B
other license
501978
#6 model · text-generation
Qwen/Qwen3.8-2.4T-A95B
other license
- What it is
- Qwen3.8-2.4T-A95B is an open-weights text generation model created by the Qwen team.
- Who it's for
- This model is for developers and researchers who have professional-grade infrastructure to host large-scale AI, as it is a 'Max-class' model designed for complex integration.
- What it does
- It excels at coding, research, and 'agentic tasks,' which are multi-step jobs where the AI plans and executes a goal autonomously. It is notable for its 'Mixture of Experts' architecture, meaning it has a massive total knowledge base but only uses a small, efficient portion of its brain for any single answer.
- Explain it like I'm five
- Imagine a giant library with 2.4 trillion books, but instead of reading every book to answer your question, the AI has a smart manager who instantly picks the 11 best experts to handle the job. This makes it incredibly powerful at solving hard problems without being too slow to use.
#7 model
Comfy-Org/MiniMax-H3
other license
1.3k6.8M
#7 model
Comfy-Org/MiniMax-H3
other license
- What it is
- MiniMax H3 is a video generation model created by MiniMax, repackaged here for use with ComfyUI, a popular node-based interface for AI tools.
- Who it's for
- This is for AI artists and developers who use ComfyUI to generate high-quality video content on their own hardware.
- What it does
- The model can create videos from text prompts (T2V), images (I2V), or reference materials (R2V). It is notable for its versatility in how it triggers video generation and its integration with the Qwen text encoder to understand complex instructions.
- Explain it like I'm five
- Imagine a digital movie studio that can turn a single sentence or a still photo into a moving scene. This specific version is like a 'plug-and-play' kit that lets people plug that studio into their own computer to make their own AI movies.
#8 model · image-to-video
lightx2v/Minimax-h3-Turbo
apache-2.0 license
41120.4k
#8 model · image-to-video
lightx2v/Minimax-h3-Turbo
apache-2.0 license
- What it is
- Minimax-h3-Turbo is an open-weights AI model developed by LightX2V based on the MiniMax-H3 architecture, designed specifically to generate video content.
- Who it's for
- This model is for developers and AI researchers who want to integrate high-quality video generation into their own applications using the Diffusers library.
- What it does
- It primarily performs image-to-video tasks, meaning it can take a still image and animate it into a moving clip. It is also capable of text-to-video and region-to-video generation, allowing for flexible ways to create motion from different types of input.
- Explain it like I'm five
- Imagine you have a photograph of a waterfall; this AI acts like a magic animator that can turn that still photo into a real moving video. It's like giving a 'play' button to a picture so the scene comes to life.
#9 model · image-text-to-text
unsloth/Muse-Glimmer-30B-GGUF
apache-2.0 license
3590
#9 model · image-text-to-text
unsloth/Muse-Glimmer-30B-GGUF
apache-2.0 license
- What it is
- Muse Glimmer is an open-weights AI model created by Meta's Superintelligence Lab that can process both text and images to perform complex tasks.
- Who it's for
- It is designed for developers and users who want to run a powerful autonomous agent locally on consumer-grade hardware, such as a home computer, without needing the cloud.
- What it does
- The model acts as an 'agent,' meaning it can plan multi-step projects, use digital tools to get things done, and fix its own mistakes if a tool fails. It is notable for its ability to 'see' screenshots and documents while reasoning through a problem in over 100 languages.
- Interesting numbers
- Strong success rates on DeepSearch QA, MCP-Atlas, 𝛕3-Bench, and SWE-Bench
- Explain it like I'm five
- Imagine a digital assistant that doesn't just talk, but can actually use your computer to finish a project. It's like giving the AI a pair of eyes to see your screen and a brain that can double-check its own work if it hits a snag, all while living entirely on your own laptop.
#10 model · image-text-to-text
moonshotai/Kimi-K3
other license
10.6k1.6M
#10 model · image-text-to-text
moonshotai/Kimi-K3
other license
- What it is
- Kimi K3 is an open-weight AI model created by Moonshot AI that can process and generate text, images, and video.
- Who it's for
- This model is designed for researchers and developers who need a high-capacity tool for complex coding, deep research, and advanced knowledge work.
- What it does
- It features a massive 2.8 trillion parameter size and a 1-million-token context window, meaning it can remember and analyze huge amounts of information at once. It is specifically built for 'agentic' work, such as managing large software repositories or designing chips with minimal human help.
- Explain it like I'm five
- Imagine a super-librarian who has read almost every book in existence and can also see and watch videos. Instead of just answering a question, this librarian can actually do the work for you—like writing a whole computer program or designing a piece of hardware—while remembering every single detail of the project.