Hugging Face Daily Papers

PAWBench: How Far Are We from Probabilistically Aligned World Modeling?

HF ★ 69 · Yuandong Pu, Le Zhuo, Sayak Paul… · HF Mirror

To address the issue that current video generation models regarded as world models only perform single-sample rationality verification and fail to cover distribution matching requirements, this paper identifies probabilistic alignment as the core indicator for world models, launches the PAWBench benchmark and PAWEval evaluation protocol to test the performance of 11 mainstream models across 50 scenarios. It finds that none of the models can cover the effective behavior range and match the reference probability at the same time, and distribution alignment can be optimized later from the dimensions of prompts, noise sampling, and training.

What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents

HF ★ 51 · Xingshan Zeng, Zishan Xu, Boju Zhang… · HF Mirror

To address the lack of a unified evaluation logic for the quality of interactive data generated by LLM Agents, this paper first disassembles agent data into a four-tuple paradigm of environment, task, interaction, and validator, and further proposes the “Accuracy-Complexity-Diversity (ACE)” analysis framework to clarify generation constraints. It points out that the field is shifting towards implementation-verified accuracy, complexity matching learner capabilities, and non-superficial diversity, with the core of producing effective non-redundant data suitable for the iteration of agents and environments.

TTPO: Test-Time Policy Optimization

HF ★ 49 · Aozhe Wang, Zhengxi Lu, Jianze Wang… · HF Mirror

To address the problem that post-training methods for LLM mathematical reasoning rely on ground truth labels and cannot perform test-time training, and majority voting pseudo-labels are prone to failure, based on the observation that “generations inconsistent with pseudo-labels are basically wrong”, this paper proposes Test-Time Policy Optimization (TTPO): it adopts an asymmetric objective, distills consistent generations, penalizes wrong examples with grouped RL, and is paired with token-level selection optimization. Without labels, its performance matches that of supervised methods, the inference accuracy of Qwen3-1.7B increases by up to 36.4%, with excellent generalization performance.

Training Agents to Evolve with Their Harness: TaoLive Digital Avatar Agent Technical Report

HF ★ 39 · TaoLive AIGC LLM Team, Yuhan Sun, Wenhao Lin… · HF Mirror

To meet the requirements of low latency and fast strategy iteration for AI digital human live streaming, this report proposes the Harness-Aware Training (HAT) method, with harness state augmentation technology as the core, and trains lightweight models in three stages to adapt to variable runtime harnesses, solving the pain points of high latency of large models and easy overfitting of small models. Its live Q&A accuracy reaches 94.8%, which is better than general large models, and the latency meets standards. After being launched on Taobao Live, both GMV and product page views have achieved positive growth.

UrbanGround: From Local Perception to Spatial Agency in a Real-Scale City

HF ★ 37 · Tianjie Ju, Zheng Wu, Yueqing Sun… · HF Mirror

To address the problem that multimodal large models can only parse static street views, and their urban spatial decision-making ability after movement remains to be verified, the team developed UrbanGround, the first real-scale urban interactive sandbox based on full 3D geographic data of Hong Kong, supporting first-person closed-loop exploration and navigation. Tests show that existing multimodal LLM agents have basic visual recognition and short-range spatial reasoning capabilities, but lack capabilities in orientation perception, obstacle avoidance, and long-term target exploration, and errors are prone to accumulate and difficult to correct.

arXiv cs.LG

SLM-Conditioned Hierarchical Relation Routing for Labeled Property Graph Learning

Michal Podstawski

To address the pain point that when conventional graph neural networks process labeled property graphs, they convert attributes into static features and cannot adapt to the dynamic selection of semantic propagation evidence for prediction targets, this paper proposes a hierarchical relation routing architecture driven by Small Language Models (SLM): it uses the output of topological GNN as structural representation anchors, uses SLM to generate target-adaptive routing queries, and performs residual updates on anchors after hierarchically filtering neighborhood messages. This method balances structural fidelity and semantic introduction, supports interpretable analysis, and provides a general paradigm for rich attribute graph learning.

NeuronFuzz: Safety Neuron Guided Fuzzing for LLM Safety Evaluation

Zhiyuan Xu, Muhammad Firhard Roslan, Joseph Gardiner…

To address the pain points that existing LLM jailbreak testing relies on response feedback, has high cost, and provides sparse guidance for strongly aligned models, this paper proposes the white-box fuzz testing framework NeuronFuzz: it extracts the activation of internal safety neurons to generate a continuous safety alert score as iterative feedback, which can run without generating model responses, combined with gradient localization of sensitive positions and masked language model generation of compliant mutation prompts. In tests on 21 models, the jailbreak detection rate reached 76-100% in white-box scenarios, up to 48 percentage points higher than the baseline, and the generated attack templates can be transferred to various models zero-shot with excellent effects.

Pruning Binarized Neural Networks: A Dedicated Framework and Globally Weighted Algorithms

Roan Rubiales, Jean Pierre David

To address the problems that existing pruning strategies have poor adaptability when binarized neural networks are adapted to edge devices and hardly bring actual hardware gains, this paper launches a dedicated research framework based on PyTorch, which supports binarized network freezing and pruning, facilitating the reproduction of existing solutions and rapid development of new methods. At the same time, a globally weighted pruning algorithm is proposed, which takes into account the importance of parameters at different levels, achieving 70% pruning rate on VGG11 without accuracy loss, far better than the 41% pruning rate of SOTA in the same scenario.

OpenAI

Supporting Thailand’s next generation of AI startups

OpenAI

This project is initiated by OpenAI in cooperation with the Ministry of Higher Education, Science, Research and Innovation of Thailand (MHESI). The core initiative is the launch of an 8-week AI special startup accelerator, focusing on three livelihood tracks: healthcare, elderly care, and education, providing targeted support for 10 local startups to help them complete the transformation from AI technology prototypes to trusted commercial products, and consolidate the development foundation of Thailand’s next-generation AI startup industry.

Better answers, broader thinking: What students gain from ChatGPT and critical-thinking training

OpenAI

The research is themed Better answers, broader thinking: What students gain from ChatGPT and critical-thinking training. It adopts a randomized trial design, includes more than a thousand college students as samples, focuses on the impact of ChatGPT use and critical thinking training on students’ real college academic performance and assignment originality, aiming to clarify the actual benefits such as improved answer quality and expanded thinking that the combination of the two can bring to students.

Anthropic News

Previewing the Model Hardware Standard

Anthropic

AI research institution Anthropic recently launched the research preview of the Model Hardware Standard (MHS), which is open for trial to the first batch of research laboratories and advanced manufacturers. MHS is a unified technical specification for AI agents, with the core function of ensuring operational safety when AI controls various physical devices. This shared standard will also provide basic support for the implementation of physical AI, industrial intelligence and other scenarios, as well as cross-entity collaboration.

How Claude’s text watermarking works

Anthropic

Anthropic announced that text generated by the Claude series of large models will be embedded with exclusive text watermarks in the future, which can quickly identify whether content is generated by Claude. This adjustment is jointly promoted by it and a number of leading AI manufacturers to meet the compliance requirements of the EU AI Act. This time, it will also publicly respond to common external concerns such as the principle of watermark technology, its impact on generation quality, and the motivation for the revision.

Google DeepMind

Gemini Omni 1.1 Flash lets you build with more control

Google DeepMind

The Gemini Omni 1.1 Flash introduced this time is a lightweight multimodal large model launched by Google, focusing on upgraded development controllability: it adds fine-grained instruction constraints and mandatory structured output functions, supports low-threshold custom fine-tuning, inference speed is 30% higher than the previous generation, call cost is reduced by 40%, balancing multimodal accuracy and cost-effectiveness, suitable for the rapid implementation of real-time interactive AI applications.

Piloting the world’s first double-blind AI evaluations

Google DeepMind

This launch of the world’s first double-blind AI evaluation pilot adopts the test rule of two-way identity concealment between evaluators and evaluated objects (AI and control humans), which can effectively avoid the interference of human preferences and targeted AI responses on evaluation results. The output AI capability conclusions are more objective and fair than traditional evaluations, and can provide more reliable evaluation basis for AI performance classification and compliance implementation.

Hugging Face Blog

Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers

Hugging Face

To address the pain point of limited representation capability of single-vector text embeddings, this paper proposes an end-to-end training and fine-tuning solution for multi-vector embedding models based on the Sentence Transformers framework, optimizing training memory usage and inference efficiency. Actual tests show that its performance on tasks such as information retrieval and semantic matching is about 10% higher than that of single-vector models, and it is compatible with the existing vector retrieval ecosystem, greatly reducing the threshold for the implementation of multi-vector embedding models.

Granite 4.2 LLMs: How They’re Built

Hugging Face

At present, only the title of this paper is provided, and the body of the abstract is missing. Please supplement the complete original abstract, and I will accurately extract its core methods, technical paths and key conclusions, and generate a clear English summary of about 120 words with highlighted key points as required.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This paper sorts out the conceptual evolution of Recursive Self-Improvement (RSI): as early as 1965, scholar I.J. Good first proposed the relevant idea, referring to superintelligent machines with intelligence far exceeding that of humans that can independently design better systems to achieve iteration. In 2008, Eliezer Yudkowsky clearly defined it as a feedback loop where AI optimizes its own cognitive mechanism with its existing intelligence. In the current AI context, this mechanism can be manifested as directly rewriting its own weights, or broadly optimizing the training process.

QbitAI

The first AGI stock on 港股 exploded! Agent business generated nearly 500 million yuan in half a year, Token revenue surged 500% in Q2

QbitAI

The first AGI stock on 港股, Unisound, released its 2026 mid-term financial report, with total revenue of 562 million yuan, a year-on-year increase of 38.7%, nearly doubling the growth rate. The core growth engine, the Agent business, recorded revenue of 478 million yuan, the newly added Token business surged more than 500% month-on-month in Q2, with a gross margin of over 60%. In contrast to the underperforming OpenClaw, Unisound has implemented Agents in scenarios such as healthcare and insurance, and divided into three revenue segments.

eCareer won the first prize in Zhejiang, advanced to the national data element competition with Ant Group and Zhejiang University

QbitAI

The Zhejiang division of the national “Data Element ×” competition led by the National Data Administration in 2026 recently concluded. The competition focuses on industrial implementation, does not only value algorithm performance, and attracted leading enterprises and universities from various fields to participate. Finally, Hailiang Tech Services Smart Education (the only first prize in the track), Ant Medical AI, and the civil aviation project jointly developed by Zhejiang University and Loong Air won the first prize in each track and advanced to the national competition.

The first cancer vaccine, 3 million yuan per shot???

QbitAI

The personalized mRNA melanoma vaccine Intismeran jointly developed by Moderna and Merck, with full participation of AI in the entire process, recently obtained positive results in phase III clinical trials. When used in combination with PD-1 inhibitors, it can delay the recurrence and metastasis of patients, and is the world’s first mRNA cancer treatment solution that has passed phase III trials. However, the vaccine is predicted to be priced at about 3.18 million yuan, and it also needs to be used in combination with the high-priced Keytruda, which may only be affordable for the rich. The pricing corresponding to the full course of treatment is still unclear.