Research tracking

Trending papers and models, explained simply

this week

#1 paper

Harness Handbook: Making Evolving Agent Harnesses Readable,Navigable, and Editable

Ruhan Wang, Yucheng Shi, Zongxia Li +7

186
What it is
A new framework designed to help developers and AI agents find and modify specific behaviors within the complex code that manages AI agents.
Who it's for
Software engineers and AI researchers who build and maintain the 'harnesses' (the surrounding code that handles prompts and tools) for AI agents.
What it does
It automatically creates a 'Harness Handbook' that maps high-level behaviors to the exact lines of code that implement them. This allows developers to quickly locate where changes need to be made, even when the logic is scattered across many different files.
Interesting numbers
  • Improved behavior localization and edit-plan quality
  • Used fewer planner tokens
  • Largest gains on scattered sites, rarely executed paths, and cross-module interactions
Explain it like I'm five
Imagine trying to fix a specific rule in a massive, 1,000-page rulebook where the instructions are scattered randomly across different chapters. The Harness Handbook is like a smart index that tells you exactly which pages to look at for a specific rule, so you don't have to hunt through the whole book manually.

#2 paper

Weak-to-Strong Generalization via Direct On-Policy Distillation

Shiyuan Feng, Huan-ang Gao, Haohan Chi +7

125
What it is
A new training method called Direct On-Policy Distillation (Direct-OPD) that transfers reasoning skills from a small AI model to a larger one.
Who it's for
AI developers and companies who want to improve large models without spending the massive amount of computing power required for full reinforcement learning.
What it does
It captures the 'lesson' a small model learned during reinforcement learning—specifically how its behavior changed—and applies that shift to a stronger model. This allows the stronger model to improve its reasoning without having to undergo the expensive trial-and-error process itself.
Interesting numbers
  • Boosted Qwen3-1.7B from 48.3% to 58.3% on AIME 2024
  • Achieved these results in just 4 hours on 8 A100 GPUs
Explain it like I'm five
Imagine a student who struggled but eventually learned a hard math trick through lots of practice. Instead of making a genius student spend weeks practicing the same way, we just show the genius exactly how the student's thinking changed from 'wrong' to 'right.' The genius then applies that specific logic shift to their own advanced knowledge to get even smarter, much faster.

#3 paper

Boogu-Image-0.1: Boosting Open-Source Unified Multimodal Understanding and Generation

Guoxuan Chen, Chufeng Xiao, Haoran Yang +30

122774
What it is
A family of open-source AI models called Boogu-Image-0.1 designed for both understanding and creating images.
Who it's for
AI developers and researchers who want high-quality image generation and editing tools without relying on expensive, closed-source corporate systems.
What it does
It allows users to generate high-quality images from text, edit existing images using instructions, and render text in both English and Chinese. It achieves these results using a more efficient training process and a smaller set of unique images than many competitors.
Interesting numbers
  • Trained using only 208.62 million unique images
  • Base model theoretical training cost of approximately $400K
Explain it like I'm five
Imagine a Swiss Army knife for images that can not only draw a picture from scratch but also look at a photo and change specific parts of it based on your request. While most 'super-brains' that do this are kept secret by big companies, this project shares the recipe and the tools for free. It proves you don't need a billion-dollar budget to make an AI that is nearly as good as the best ones in the world.

#4 paper

VideoChat3: Fully Open Video MLLM for Efficient and Generalist Video Understanding

Xinhao Li, Yuhan Zhu, Xiangyu Zeng +24

10743
What it is
VideoChat3 is a new, fully open-source Multimodal Large Language Model (MLLM), which is an AI capable of understanding and discussing both text and video.
Who it's for
AI researchers and developers who want a high-performing video AI that is efficient to run and completely transparent in how it was built.
What it does
It improves how AI processes video by using a more efficient way to 'see' frames and a smart system that adjusts image resolution for streaming video. It also introduces three new high-quality datasets to help the AI better understand general, long-form, and streaming videos.
Interesting numbers
  • Surpasses prior open-source models with equal or larger parameter counts using only 4B parameters
Explain it like I'm five
Imagine trying to watch a movie but only seeing one still photo every few minutes; you'd miss a lot. VideoChat3 is like giving the AI a better way to 'watch' the flow of a video without needing a supercomputer to do it. It's also 'fully open,' meaning the creators shared the secret recipe so anyone can learn from it or improve it.

#5 paper

Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models

Yubo Wang, Jiarong Liang, Yuxuan Zhang +5

9512
What it is
A new training method called 'function-aware fill-in-the-middle' (FIM) designed to improve how AI coding agents handle external tools.
Who it's for
Researchers and companies building AI coding agents that need to interact with real-world software environments.
What it does
It teaches AI models to better understand the 'loop' of calling a function and using its result to continue a task. By masking specific parts of code during training, it helps the model integrate external tool returns into its reasoning process more effectively.
Interesting numbers
  • SWE-Bench-Verified improved by +2.8/+3.0 for 7B/14B models
  • SWE-Bench-Lite gains of +3.7/+4.0/+5.4 across models
  • Mid-training used a 2.6B-token corpus from 968 GitHub repositories
Explain it like I'm five
Imagine an AI learning to cook by reading recipes. Usually, it just reads from top to bottom, but it struggles when a recipe says 'wait for the oven to beep' before continuing. This method teaches the AI to 'fill in the blanks' around those waiting periods, making it much better at using tools and reacting to what happens in the real world.

#6 paper

ABot-N1: Toward a General Visual Language Navigation Foundation Model

Ruiyan Gong, Yingnan Guo, Junjun Hu +43

95
What it is
ABot-N1 is a new foundation model designed for Visual Language Navigation, which is the ability for a robot to move through an environment based on visual cues and language instructions.
Who it's for
Researchers and companies building autonomous robots, such as delivery bots or indoor assistants, that need to navigate complex real-world environments.
What it does
It separates 'thinking' from 'moving' by using a slow reasoner to pick a visual target and a fast expert to handle the actual steering. This prevents the robot from getting lost or confused by rare objects and makes its decision-making process easier for humans to understand.
Interesting numbers
  • Boosted POI (Point of Interest) arrival by 35.0% to 77.3%
  • Achieved 95.4%/92.9% SR (Success Rate) in complex indoor and outdoor scenes
Explain it like I'm five
Imagine a robot that works like a human with a map: first, it slowly thinks and puts a 'pin' on a picture of where it wants to go (the slow reasoner), and then it quickly steers itself toward that pin (the fast expert). This is better than older robots that tried to do both at once and often got confused. It's like having a navigator and a driver working together instead of one person trying to do both while blindfolded.

#7 paper

Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning

Xinyu Tang, Gangqiang Cao, Yurou Liu +13

84
What it is
A research project that scales 'Zero RL'—a training method where AI learns to reason by testing its own answers against a verifiable truth rather than following human examples—to a massive model with one trillion parameters.
Who it's for
AI researchers and companies building large-scale reasoning models that want to understand how AI behavior changes as models get bigger.
What it does
It introduces a stable training pipeline that prevents the AI from becoming repetitive or incoherent as it grows. This enables the model to spontaneously develop advanced habits like self-checking its work and organizing its thoughts without being told to do so by humans.
Interesting numbers
  • Evaluated on seven mathematical benchmarks
  • Competitive performance with Ring-2.5-1T-Zero
Explain it like I'm five
Imagine a student who learns math not by reading a textbook, but by guessing answers and only being told 'Correct' or 'Incorrect.' This paper shows that if you give that student a massive 'brain' (a trillion parameters), they don't just get the right answers—they actually start teaching themselves how to show their work and double-check for mistakes automatically.

#8 paper

ABot-AgentOS: A General Robotic Agent OS with Lifelong Multi-modal Memory

Jiayi Tian, Shiao Liu, Yuting Xu +30

78
What it is
A new software framework called ABot-AgentOS that acts as an 'operating system' for robots, providing a layer for reasoning and memory on top of their physical controllers.
Who it's for
Researchers and developers building advanced robots that need to perform complex, long-term tasks in diverse environments.
What it does
It allows robots to remember visual and spatial information over time using a multi-modal graph memory. It also includes a self-evolution loop that lets the robot learn from its own failures to improve future performance.
Interesting numbers
  • 87.5 on LoCoMo (improved to 88.7 with self-evolution)
  • 59.9 on OpenEQA EM-EQA (improved to 60.4 with self-evolution)
  • 88.6 on Mem-Gallery (improved to 89.0 with self-evolution)
  • 76.5 Acc@All on NExT-QA
Explain it like I'm five
Imagine a robot that has 'muscles' to move but no 'brain' to remember where it left your keys. This project gives the robot a digital diary and a planning system so it can remember things it saw and learn from its mistakes, much like how a computer's OS manages apps and files.

#9 paper

Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation

Haozhe Wang, Weijia Feng, Jinpeng Yu +8

7836
What it is
A new framework and dataset designed to help AI image generators use search tools to find information they weren't taught during training.
Who it's for
AI researchers and developers building image generation tools that need to accurately depict real-world people, current events, or niche objects.
What it does
It introduces a 'teach-then-search' training method that helps the AI recognize the boundary between what it already knows and what it needs to look up. This prevents the AI from hallucinating (making up fake details) or adding unnecessary search noise to simple requests.
Interesting numbers
  • Frontier open generators scored only 21 to 28 out of 100 on the new SearchGen-Bench
  • A 40-point collapse in performance that was invisible to previous benchmarks
Explain it like I'm five
Imagine an artist who is great at painting but doesn't know what a brand-new gadget looks like. Instead of just guessing and painting something fake, this system teaches the artist exactly when to stop and use Google to find a reference photo before they start painting.

#10 paper

Video Generation Models are General-Purpose Vision Learners

Letian Wang, Chuhan Zhang, Rishabh Kabra +9

74
What it is
A new framework called GenCeption that turns a video-generation AI into a tool for understanding and analyzing visual data.
Who it's for
AI researchers and developers building computer vision systems for robotics, 3D mapping, and video analysis.
What it does
It uses a model originally trained to create videos to instead perform complex perception tasks, such as estimating depth, tracking 3D keypoints, and identifying objects. It proves that learning how to generate a video helps an AI understand the physical laws and geometry of the real world.
Interesting numbers
  • Achieves comparable performance with leading models using 7 to 500 times less training data
  • Matches or surpasses specialized models like DepthAnything3 and SAM3
Explain it like I'm five
Imagine if you learned how to paint a perfect 3D sculpture from scratch; you'd suddenly be an expert at spotting mistakes in other people's sculptures. This research shows that an AI trained to 'paint' realistic videos becomes an expert at 'seeing' and understanding the 3D world, even for things it has never seen before.
on the Hugging Face hub

#1 model · image-text-to-text

thinkingmachines/Inkling

apache-2.0 license

9357.9k
What it is
Inkling is an open-weights multimodal AI model created by Thinking Machines that can process text, images, and audio to generate text responses.
Who it's for
It is designed for developers building AI applications like coding assistants, chatbots, and agentic systems (AI that can use tools to complete tasks).
What it does
Inkling can 'see' images and 'hear' audio files to answer questions or follow instructions in multiple languages. It is notable for using a Mixture-of-Experts (MoE) architecture, which means it has a massive total capacity but only activates a small fraction of its parameters for any given task to stay efficient.
Explain it like I'm five
Imagine a giant library where 256 specialized librarians live. Instead of every librarian trying to answer every question, Inkling only wakes up the 8 best experts for the specific job at hand. This allows it to be a 'jack-of-all-trades' that can understand a photo, a voice recording, or a line of code all at once.

#2 model · text-generation

prism-ml/Ternary-Bonsai-27B-gguf

apache-2.0 license

666200.8k
What it is
Ternary Bonsai 27B is a text-generation AI model created by Prism ML, based on the Qwen3.6-27B architecture.
Who it's for
It is designed for people who want to run a powerful AI on consumer hardware, such as a standard laptop or a single GPU, including Apple Silicon devices.
What it does
The model uses a 'ternary' system to shrink its size significantly without losing much intelligence, allowing it to handle complex reasoning, coding, and math on-device. It also supports a very large context window of 262K tokens, meaning it can process long documents without running out of memory.
Interesting numbers
  • ~9.4x smaller than FP16 (ideal)
  • 95% of FP16 intelligence retained
  • ~26 tok/s on an Apple M5 Pro laptop
  • ~7.2 GB deployed footprint (down from ~54 GB FP16)
Explain it like I'm five
Imagine trying to fit a giant encyclopedia into a tiny pocket notebook. Usually, if you shrink the text too much, it becomes unreadable gibberish. Prism ML found a way to 'compress' the AI's brain so it fits on a laptop, but it still remembers almost everything and can still solve hard math problems.

#3 model · text-generation

prism-ml/Bonsai-27B-gguf

apache-2.0 license

3811.0M
What it is
Bonsai-27B is a text-generation AI model created by Prism ML, based on the Qwen3.6-27B architecture.
Who it's for
It is designed for people who want to run a powerful AI on everyday hardware, such as laptops (Apple M5 Pro), single GPUs, or even iPhones.
What it does
The model uses '1-bit' weights, which drastically shrinks its size so it can fit on small devices without losing most of its intelligence. It is capable of complex reasoning, coding, and math, and can handle very long documents (up to 262K tokens) thanks to a specialized hybrid-attention system.
Interesting numbers
  • ~14.2x smaller than FP16
  • ~90% of FP16 intelligence retained
  • ~44 tok/s on an Apple M5 Pro laptop
  • ~3.9 GB deployed footprint (down from ~54 GB FP16)
Explain it like I'm five
Imagine taking a massive, heavy encyclopedia and finding a way to shrink it into a tiny pocket notebook without losing the important information. Bonsai does this with AI 'weights,' making a giant model small enough to run on your phone while still being smart enough to help you with homework or coding.

#4 model · image-text-to-text

empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF

apache-2.0 license

2.3k2.1M
What it is
Qwythos-9B is an open-weights multimodal AI model developed by Empero, based on the Qwen3.5 architecture that can process both text and images.
Who it's for
It is designed for developers and AI enthusiasts who want a high-reasoning model that can run locally on consumer hardware, such as a laptop, using tools like llama.cpp or Ollama.
What it does
The model specializes in complex reasoning and 'chain-of-thought' processing, with a massive 1-million-token context window for analyzing very long documents. It also features native function calling and vision capabilities, making it useful for technical tasks in fields like cybersecurity and biomedicine.
Interesting numbers
  • +34 pts MMLU
  • +30 pts gsm8k-strict
  • +19 pts gsm8k-flex
Explain it like I'm five
Imagine a student who didn't just memorize facts but learned exactly how to 'think through' hard problems step-by-step. This model is like that student, but with a photographic memory that can hold a whole library of books in its head at once while also being able to 'see' and describe images.

#5 model · text-generation

zai-org/GLM-5.2

mit license

4.1k534.7k
What it is
GLM-5.2 is a text-generation AI model created by z.ai that is released under an open-source MIT license.
Who it's for
It is designed for developers and researchers who need a powerful, open-weights model for complex coding and long-document analysis.
What it does
The model is built for 'long-horizon tasks,' meaning it can remember and process a massive amount of information (up to 1 million tokens) at once. It also features flexible 'thinking effort' levels for coding, allowing users to trade off speed for higher quality results.
Interesting numbers
  • AIME 2026: 99.2
  • FrontierSWE (Dominance): 74.4
  • IMOAnswerBench: 91.0
  • DeepSWE: 46.2
Explain it like I'm five
Imagine an AI with a giant digital whiteboard that can hold a million pieces of information without forgetting the beginning. It's like a super-student who can read a whole library of books and then write a complex computer program based on everything it read. Plus, you can tell it to 'think fast' for easy tasks or 'think deep' for the hard ones.

#6 model · image-text-to-text

bottlecapai/ThinkingCap-Qwen3.6-27B

apache-2.0 license

4099.4k
What it is
ThinkingCap is a specialized version of the Qwen3.6-27B model created by BottleCap AI. It is an image-text-to-text model, meaning it can process both visual images and written text to generate responses.
Who it's for
This model is for developers and AI researchers who want a high-reasoning model that is faster and cheaper to run by using fewer computational resources.
What it does
It provides the same high-quality reasoning and answer style as the original Qwen model but does so more efficiently. It is notable for significantly reducing 'thinking tokens'—the internal processing steps the AI takes before giving a final answer—without sacrificing accuracy.
Interesting numbers
  • 50% less thinking tokens on average
  • over 90% less [thinking tokens] in best cases
Explain it like I'm five
Imagine a genius student who used to write ten pages of scratchpad notes before answering a math problem. ThinkingCap is like that same student, but they've learned how to get the right answer using only five pages of notes. It's just as smart, but it gets to the point much faster.

#7 model

conradlocke/krea2-identity-edit

other license

3410
What it is
This is an unofficial community-made add-on (called a LoRA) for the Krea 2 image model, designed to perform precise image editing.
Who it's for
It is for digital artists and AI hobbyists who use ComfyUI (a node-based interface for AI generation) and have the necessary hardware to run large image models.
What it does
The model allows users to edit images using plain-language instructions while keeping a person's identity—like their face and moles—exactly the same. It can move a person to a new scene, swap faces, or change outfits without altering the rest of the image.
Explain it like I'm five
Imagine having a magic eraser and a paintbrush that knows exactly who a person is. You can tell the AI, 'Put this specific person at a night market,' and it will move them to the new location while keeping their face and clothes perfectly identical, rather than creating a random stranger who looks similar.

#8 model · image-text-to-text

HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

apache-2.0 license

2.8k2.3M
What it is
This is a multimodal AI model created by HauhauCS, based on the Qwen3.6 architecture, which means it can process and respond to both text and images.
Who it's for
It is for advanced users who want an AI without safety restrictions, provided they have a computer with enough RAM (ranging from 15GB to 44GB depending on the version chosen) to run the GGUF files, which are a format designed for efficient local playback.
What it does
The model is designed to be 'uncensored,' meaning it is modified to remove the refusals typically found in standard AI assistants. It uses a Mixture of Experts (MoE) design—where only a fraction of the model's parameters are used for each task to increase efficiency—and supports multiple languages including English and Chinese.
Interesting numbers
  • 0/465 Refusals
Explain it like I'm five
Imagine a super-smart digital assistant that can 'see' photos and chat in different languages, but usually has a strict set of rules about what it can't talk about. This version is like the same assistant but with the rulebook ripped out, so it will answer almost any question you throw at it without saying 'I can't do that.'

#9 model · audio-text-to-text

OpenMOSS-Team/MOSS-Transcribe-Diarize

apache-2.0 license

24783.2k
What it is
MOSS-Transcribe-Diarize is an open-weights audio understanding model created by the OpenMOSS Team that converts speech into text.
Who it's for
Developers and researchers who need to transcribe long recordings, such as podcasts or business meetings, and want a model they can run on their own infrastructure.
What it does
It performs transcription and 'diarization'—the process of identifying who is speaking—simultaneously in one pass. It can handle audio up to 90 minutes long across 50+ languages and can be prompted with 'hotwords' to help it recognize specific technical terms.
Interesting numbers
  • Won first place in the 2nd MLC-SLM Challenge at INTERSPEECH 2026
Explain it like I'm five
Imagine a court reporter who not only types every word spoken in a room but also automatically labels who said what (like 'Speaker 1' or 'Speaker 2') and notes exactly when they spoke. Instead of using three different tools to do this, this AI does it all at once, even for very long recordings.

#10 model · image-text-to-text

ATH-MaaS/OvisOCR2

apache-2.0 license

15010.8k
What it is
OvisOCR2 is a compact, multimodal AI model designed for document parsing, built by fine-tuning the Qwen3.5-0.8B base model.
Who it's for
It is for developers and businesses that need to convert images of documents into digital text; because it is a small 0.8B parameter model, it has a small deployment footprint suitable for efficient hosting.
What it does
The model reads an image of a page and converts it into a Markdown document, preserving the natural reading order of text, tables, and mathematical formulas. It is notable for being the first 'end-to-end' model (meaning it does the whole job in one step rather than using a chain of different tools) to top the OmniDocBench v1.6 leaderboard.
Interesting numbers
  • Overall score of 96.58 on OmniDocBench v1.6
  • Highest Avg3 score of 75.06 on PureDocBench
Explain it like I'm five
Imagine a super-smart scanner that doesn't just take a picture of a page, but actually understands where the tables, math equations, and paragraphs are. Instead of just giving you a flat image, it rewrites the whole page into a clean digital document that looks exactly like the original. It's like having a professional typist who can read any messy page and instantly turn it into a perfectly formatted digital file.