Hugging Face Daily Papers

Cosmos 3: Omnimodal World Models for Physical AI

Authors: Aditi, Niket Agarwal, Arslan Ali, Jon Allen, Martin Antolini…
HF Votes: 17
Hugging Face: https://huggingface.co/papers/2606.02800

AI Summary:

NVIDIA has launched Cosmos 3, an omnimodal world model for physical AI. It adopts a unified hybrid Transformer architecture that supports flexible input and output configurations, enabling joint processing and generation of text, images, audio, video, and action sequences to break cross-modal capability barriers. The model achieves new SOTA performance across multiple understanding and generation tasks, serving as a general-purpose backbone for embodied intelligence. Its related sub-models rank first in multiple open-source tracks, and all supporting development resources have been open-sourced.

Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories

Authors: Jiaming Wang, Ziteng Feng, Jiangtao Wu, Ruihao Li, Qianqian Xie…
HF Votes: 12
Hugging Face: https://huggingface.co/papers/2606.02060

AI Summary:

To address the pain point that deep research agents cannot locate trajectory errors when evaluated solely on final results, the research team collected 2,790 real running trajectories, annotated them, and built TELBench, an error localization benchmark with 1,000 instances. They also proposed the claim-centric DRIFT auditing framework, which locates errors by matching claims with trajectory evidence. Experiments show that DRIFT can improve error localization and first-error detection accuracy by up to 30 percentage points, providing a process-focused perspective for agent reliability research.

Echo-Infinity: Learning Evolving Memory for Real-Time Infinite Video Generation

Authors: Yuxuan Bian, Zeyue Xue, Songchun Zhang, Shiyi Zhang, Weiyang Jin…
HF Votes: 10
Hugging Face: https://huggingface.co/papers/2606.04527

AI Summary:

This paper proposes Echo-Infinity, an autoregressive framework for real-time infinite video generation. First, drawing on the human memory consolidation mechanism, it replaces manual KV cache strategies with learnable evolving memory, processing history of any length with constant overhead. Second, it designs a unified relative RoPE solution that breaks pre-trained length limits and eliminates training-inference bias. The solution achieves SOTA performance, realizing 24-hour real-time generation of over one million frames for the first time, providing a feasible path for infinite video generation.

ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning

Authors: Ziyan Liu, Xueda Shen, Yuzhe Gu, Songyang Gao, Kuikun Liu…
HF Votes: 10
Hugging Face: https://huggingface.co/papers/2606.03503

AI Summary:

To address the flaw that result-oriented chain-of-thought reinforcement learning for large reasoning models tends to reinforce redundant exploration and cause overthinking, this paper proposes the ThoughtFold framework. It uses an introspective strategy to identify redundant candidate sub-trajectories in correct reasoning trajectories, then applies masked preference optimization to penalize redundancy and splice core reasoning segments to fold and compress the reasoning chain. Experiments show that it reduces inference token consumption of the target 7B model by 56% while maintaining optimal accuracy.

Authors: Jie Huang, Ruixun Liu, Sirui Sun, Xinyi Yang, Yin Li…
HF Votes: 7
Hugging Face: https://huggingface.co/papers/2606.05008

AI Summary:

To fill the gap of systematic memory capability evaluation in the current multimodal long-video understanding field, this paper proposes M³Eval, the first multimodal memory evaluation benchmark based on cognitive psychology, with dedicated tasks designed to isolate different memory dimensions. Tests on mainstream models found that they generally suffer from issues such as easy entanglement of parallel video stream representations, memory interference patterns different from humans, weak temporal traceability, and insufficient symbolic memory, providing support for subsequent optimization of model memory mechanisms.

arXiv cs.LG (Machine Learning)

Early Detection of Alzheimer’s Disease Using Explainable Machine Learning on Clinical Biomarkers: A Multi-Class Classification Study Using the Alzheimer’s Disease Neuroimaging Initiative (ADNI) Dataset

Author: Afshan Hashmi

AI Summary:

This study targets the demand for early screening of Alzheimer’s disease (AD). Based on 8 routine clinical features from the ADNI dataset, it builds an explainable XGBoost three-class classification model to distinguish normal cognition, mild cognitive impairment, and AD. It uses SMOTE to handle class imbalance and SHAP for feature interpretation. The model achieves an excellent macro AUC of 0.982 on the test set, identifies core predictive features for different categories, and enables high-precision, clinically reasonable AD screening using only routine clinical indicators.

Novel Aspects of IEEE SA P3109 Arithmetic Formats for Machine Learning

Authors: Andrew Fitzgibbon, Christoph M. Wintersteiger, Jeffrey Sarnoff

AI Summary:

This paper outlines the new features of the IEEE P3109 floating-point standard draft for machine learning. It defines a family of parameterizable binary floating-point formats adapted for efficient numerical representation in low-bit scenarios; it supports exception-free operations, multiple rounding and saturation modes including built-in stochastic rounding, unifies operation rules for co-scaling factor blocks, and adds the kappa approximation metric to standardize vendors’ approximate implementations. All standard definitions have been formally verified.

Position: Deployed Reinforcement Learning should be Continual

Authors: Parnian Behdin, Kevin Roice, Golnaz Mesbahi

AI Summary:

This is a position paper on reinforcement learning (RL), pointing out that currently deployed RL generally adopts the “train first, then freeze” paradigm, only restarting training when performance degrades, which has obvious limitations. The core claim of the paper is that deployed RL with evaluative reward signals is essentially a continual learning problem, and four types of non-stationary factors after deployment require continuous model adaptation. Combined with real successful cases, it also presents the advantages of replacing the old paradigm and the path forward for adoption.

OpenAI Official News

Introducing new capabilities to GPT-Rosalind

Author: OpenAI

AI Summary:

This upgrade targets GPT-Rosalind, the dedicated large model for the life sciences field, focusing on expanding four core capabilities: improved reasoning accuracy for professional biology questions, full-chain professional knowledge addition for the medicinal chemistry field, new deep genomics data analysis functions, and support for full-process experimental solution design and implementation. The upgraded model covers multiple scenarios including basic research and new drug development, effectively improving the quality and efficiency of life sciences research.

How Wasmer used Codex to build a Node.js runtime for the edge

Author: OpenAI

AI Summary:

This article introduces the development practice of WebAssembly runtime service provider Wasmer: the team used the Codex large model auxiliary tool powered by GPT-5.5 to develop a Node.js runtime for edge scenarios, increasing development efficiency by 10 to 20 times compared to the traditional model. The original R&D delivery cycle that took months was shortened to weeks, verifying the significant efficiency improvement effect of large models on underlying toolchain development.

Anthropic News

Introducing Claude Opus 4.8

Author: Anthropic

AI Summary:

Claude Opus 4.8 is the latest upgraded version of the Opus series large models for high-end application scenarios. This version has significantly improved performance in three core scenarios: coding development, agent tasks, and professional work across various fields. It also greatly enhances the processing stability and output consistency for long-cycle complex tasks, better adapting to the needs of high-difficulty, long-span continuous work.

Introducing Claude Design by Anthropic Labs

Author: Anthropic

AI Summary:

Anthropic Labs has newly released Claude Design, an AI productivity tool for creation scenarios, which supports users to collaborate with the Claude large model to complete professional visual content creation. It can output various high-quality visual works including design drafts, interactive prototypes, presentation slides, and single-page promotional materials. The tool does not require users to master professional design skills, greatly lowering the threshold for creating high-quality visual content, and adapts to the needs of multiple office and presentation scenarios.

Google DeepMind

We’re launching the Google DeepMind Accelerator program in Asia Pacific to tackle environmental risks

Author: Google DeepMind

AI Summary:

Google DeepMind recently launched a special accelerator program in the Asia-Pacific region, positioned to address various regional environmental risks. Relying on its leading AI technology accumulation, the program provides support such as technical assistance and resource matching for environmental-focused tech innovation teams and research entities in the Asia-Pacific region, promoting the application of AI in scenarios including climate modeling, disaster early warning, and ecological protection, and improving regional environmental risk response efficiency.

Fast-tracking genetic leads to reverse cellular aging

Author: Google DeepMind

AI Summary:

This study targets the demand for identifying genetic targets to reverse cellular aging. Its core method involves biologists carrying out screening work with the help of the Co-Scientist tool, ultimately successfully discovering a batch of previously unreported novel regulatory factors. Verification confirms that these factors can effectively achieve rejuvenation reprogramming of human cells, providing brand-new research directions and candidate targets for subsequent anti-aging technology R&D and intervention for aging-related diseases.

Hugging Face Blog

Direct Preference Optimization Beyond Chatbots

Author: Hugging Face

AI Summary:

Currently, only the title of this post Direct Preference Optimization Beyond Chatbots has been provided, with no specific summary content attached. Please supplement the full original text of the summary, and I will extract the core methods and conclusions as required to output a concise, clear Chinese summary of approximately 120 words.

Adding MCP Tools to Reachy Mini

Author: Hugging Face

AI Summary:

Currently, only the title of this post has been provided, and the specific content of the summary has not been pasted. Please supplement the full English original of this post’s summary, and I will extract the core methods and conclusions as required to output a concise, accurate Chinese summary of around 120 words with clear information.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

Author: The Gradient

AI Summary:

This paper studies AI alignment from the perspective of virtue ethics, refuting the premise that “rational agents are necessarily oriented towards fixed goals”. It proposes that human rationality essentially adapts actions to a self-driven practice network that includes norms, evaluation standards, and supporting resources. To realize AI that adapts to human ethical requirements, meets core security attributes, and supports human-machine collaboration, AI decision-making logic needs to match this practice-based action logic of humans.

QbitAI

LeCun 10亿押注的方向，全球领先视觉大模型团队早已布局

Author: QbitAI

AI Summary:

The Shenzhen-based Shiqi Future team, which launched the world’s first vision large model Grounding DINO, has already made early arrangements for the world model track that Yann LeCun has bet $1 billion on, focusing on the more difficult latent space world model: learning the causal laws between actions and world states in the abstract representation space, solving the pain points of high interaction cost and low sample efficiency for agents in the physical world, and supporting agents to realize predictive decision-making.

一个GPT Plus会员的钱，够机器人跑一个月世界模型了

Author: QbitAI

AI Summary:

Zhizai Wujie has released Being-H-Flash, the world’s first implicit world model that can run on 100 TOPS-level edge chips. In the scenario of a robot sorting 1,000 express packages per day, its monthly computing power cost is only 150 yuan, equivalent to the price of a GPT Plus subscription, only 2% of NVIDIA’s Cosmos solution, and 70% cheaper than the VLA architecture Pi0.5. It can achieve near 20FPS real-time operation on the Orin NX edge chip, is compatible with NVIDIA and domestic Chinese AI chips, and does not require cloud deployment.

戴盟机器人完成亿元融资，阿里通义多模态大牛加盟攻关物理世界模型

Author: QbitAI

AI Summary:

Embodied intelligence enterprise Dimeng Robotics recently completed a 100 million yuan Series A financing, jointly invested by Inovance Capital and China Telecom. Yuan Weihao, former multimodal expert at Alibaba Tongyi, has joined as Chief AI Scientist. The team focuses on the tactile intelligence route, focusing on tackling physical world models that integrate touch and contact state. The financing will be used for model R&D, construction of related datasets, and commercial closed-loop implementation.