Hugging Face Daily Papers

FrontierChallenge: Evaluating Scientific Workflow Completion

HF ★ 94 · Liangcai Su, Zhaopeng Feng, Zhuo Chen… · HF Mirror

To address the limitation that existing scientific agent benchmarks mostly focus on single-domain, isolated tasks, the research team launched FrontierChallenge, a cross-domain end-to-end scientific workflow evaluation benchmark, which currently has 97 evaluation tasks covering 6 scientific fields. Tests on 12 cutting-edge models paired with 3 agent architectures show that the highest pass rate is only 20.6%, and high process scores or self-reported completion by models do not mean the task is actually qualified, highlighting the necessity of full-process delivery integrity evaluation.

VoiceMem: Streaming Dual-Brain Memory for Real-Time Interaction

HF ★ 68 · Zhifei Xie, Jiaqi Lang, Ze An… · HF Mirror

To solve the problem that dialogue systems lack high-performance streaming memory systems suitable for real-time interaction, this paper proposes VoiceMem, a dual-brain memory architecture divided into a left brain for information processing and a right brain for emotion and persona processing, paired with a streaming IO mechanism and complete training, evaluation, and pluggable deployment pipelines. Experiments show that its retrieval accuracy is nearly 30 points higher than Mem0, its emotion and persona performance improves SOTA by 4.29 points, and retrieval only takes 134ms with no extra latency, which can support real-time personalized emotional voice interaction.

WarpSAC: Towards the Pinnacle of Scalable Off-policy RL by Rethinking Exploration and Exploitation

HF ★ 68 · Zihao Wu, Hongyao Tang, Yi Ma… · HF Mirror

This study finds that the effect of existing stabilizers for off-policy reinforcement learning varies with data scale, and accordingly proposes the scenario-aware WarpSAC algorithm family: enable parameter normalization and dual-Q clipping for CPU scenarios with limited data, disable normalization and switch to single-Q for GPU scenarios with sufficient data, paired with sample weight decay to improve efficiency. Compared to the baseline FlashSAC, its multi-scenario score, operation success rate, and real-machine migration speed are all significantly improved, verifying the necessity of adapting stabilizers to data scenarios.

VGI-BENCH: Probing Visual Intelligence in Video Generation Models

HF ★ 46 · Xuan He, Cong Wei, Yuhao Cheng… · HF Mirror

To address the lack of highly adaptable, reasonably difficult and reliable evaluation for the zero-shot visual reasoning ability of existing video generation models, this research launched the VGI-Bench benchmark, which includes 810 instances across 27 tasks, and adopts a two-level classification of task domains and skill tags to achieve fine-grained evaluation. Actual tests show that current models perform poorly: the strongest model Seedance 2.0 only has an accuracy rate of 51%, with weak reasoning and error correction ability. The benchmark will be open-sourced to promote research and development in the field.

JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution

HF ★ 36 · Guibin Zhang, Leo Lu, Fangzhou Xie… · HF Mirror

To solve the pain points of previous agent orchestration frameworks that rely on manual design, have poor task adaptability and are difficult to scale, this paper proposes JIT-Agent: the first specially trained just-in-time orchestration intelligent model, which can generate composable frameworks adapted to tasks on demand, and supports runtime repair, experience distillation and self-evolution. Tests show that the performance of multiple large models such as DeepSeek and GLM integrated with this framework can be improved by up to 20.2 points, on par with mature commercial orchestration solutions, confirming that orchestration intelligence is a new dimension of agent capability independent of model scaling.

arXiv cs.LG

Dynamic Influence-Weighted Distillation for Single-IMU Activity Recognition

Bingxuan Xie

To address the insufficient accuracy of single right-arm IMU human activity recognition, this paper proposes a dynamic influence-weighted knowledge distillation method: during training, the logits and features of the multi-IMU teacher model are used as supervision signals, and the effect of single-step updates is verified through the internal validation split of the training set, to dynamically assign weights to the two types of loss for each sample. On the WEAR dataset, this method improves the macro F1 score by 7.66 and 6.68 percentage points compared to the supervised baseline and fixed-weight distillation respectively, and deployment does not require changes to the original sensing scheme and student model structure.

GreenLeaf Law Embed Tiny: A Compact Embedding Model for Legal Domain Retrieval

Surya Saka

This paper launches GreenLeaf Law Embed Tiny, a lightweight legal domain retrieval embedding model with 0.6B parameters. It adopts two-stage training: first distill knowledge from large models, then perform domain fine-tuning with hard negative mining combined with 3.4 million legal data entries from multiple jurisdictions, including 150,000 manually annotated samples, supporting multi-precision quantized inference. It achieves 75.11% and 64.38% on two legal retrieval benchmarks respectively, leading among models with less than 1B parameters, and is suitable for resource-constrained deployment scenarios.

Multi-Modal Anomaly Detection: A Survey

Xudong Mou, Zexin Wu, Chuan Luo…

This survey on multi-modal anomaly detection addresses the shortcomings of scattered existing research and previous surveys that mostly classify by architecture, and sorts out work in the field from a hypothesis-driven perspective: methods are divided into two categories, one based on the normality hypothesis, which models normal patterns through representation learning, cross-modal alignment, etc.; the other based on the anomaly hypothesis, which sharpens decision boundaries by injecting various types of anomalies. In addition, it sorts out the innovations brought by large models to this field, and provides benchmarks and future research directions.

OpenAI

Bringing ChatGPT for Teachers to more U.S. school districts

OpenAI

The education-scenario customized ChatGPT for Teachers is expanding its rollout in the US, and has now added coverage of 55 public school districts, providing AI functions that meet campus data security standards for more than 100,000 frontline teachers and administrative staff, along with exclusive usage training and follow-up technical support, to help education practitioners improve teaching and school administration efficiency with AI.

Learning never stops: How AI makes learning continuous

OpenAI

This research focuses on the mechanism of AI enabling continuous learning, with core reference to the latest special report released by OpenAI, sorting out and analyzing ChatGPT usage scenarios and actual effectiveness for teachers and students. The conclusion shows that ChatGPT can break the time and space limitations of traditional classrooms, provide supporting learning support outside the classroom, effectively support the continuity of the learning process, and provide a feasible direction for the implementation of full-scenario regular learning.

Anthropic News

Introducing Claude Opus 5

Anthropic

Claude Opus 5 is a major iterative version of Anthropic’s Opus-tier flagship large model, which achieves two core performance upgrades: first, its support for long-running agents has achieved a stepwise improvement, which can adapt to the deployment needs of long-process agents; second, it has optimized the performance of code generation and professional domain task processing, which can further meet the efficient operation needs of complex programming and various professional scenarios.

How Claude’s text watermarking works

Anthropic

Anthropic announced that all future Claude series large models will have built-in text watermarking function, which can quantitatively determine the probability that the text to be detected is generated by Claude. This measure is part of its joint effort with multiple leading AI vendors to implement the compliance requirements of the EU AI Act. This disclosure also addresses public concerns about the principle of watermarking technology, its impact on model output quality, the motivation for its launch and other related issues.

Google DeepMind

Intelligent transcription with Gemini 3.5 Transcribe

Google DeepMind

The newly launched Gemini 3.5 Transcribe is a new generation of intelligent speech-to-text tool, with its core upgrade being higher-level intelligent transcription capability. Compared with traditional similar products, its transcription accuracy and adaptability to complex scenarios are significantly improved, which can efficiently complete speech content transcription in multiple contexts, effectively reduce manual proofreading costs, and meet the transcription needs of multiple scenarios such as office, academia, and content creation.

From Atari to EVE Online: Building on 15 Years of AI Research in Games

Google DeepMind

This article sorts out Google DeepMind’s 15 years of progress in game AI research: early on it used lightweight games such as Atari as test beds, verifying the feasibility of reinforcement learning technology implementation; now through in-depth cooperation with game studios, it develops breakthrough AI gameplay prototypes for highly complex open-world games such as EVE Online, fully confirming the core supporting role of game scenarios for the iteration of general artificial intelligence.

Hugging Face Blog

Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers

Hugging Face

Only the title of the paper is currently available, and the full text of the abstract is missing. Please provide the complete English abstract of this paper, and I will translate and extract the core information for you, and produce a concise Chinese summary of around 120 words that highlights the method and conclusion as required.

Granite 4.2 LLMs: How They’re Built

Hugging Face

Only the title of the paper is currently available, and the full text of the abstract is missing. Please provide the complete abstract text of Granite 4.2 LLMs: How They’re Built, and I will extract it into a Chinese content of around 120 words as required, which clearly highlights the model’s construction method, core features and relevant conclusions.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

The concept of Recursive Self-Improvement (RSI) originated from the “ultraintelligent machine” setting proposed by I.J. Good in 1965: this type of system has intelligence far exceeding that of humans, and can design better machines by itself to complete self-iteration. In 2008, Eliezer Yudkowsky explicitly defined it as a feedback loop where AI uses its own intelligence to optimize its underlying cognitive mechanism. In the current AI context, this feedback loop can be manifested as the model directly rewriting its own weights, or broadly cover the behavior of the model optimizing its own training pipeline.

QbitAI

Jiyuan Lüdong raised tens of millions of dollars in cumulative financing, launched the “Chinese version of OpenRouter”

QbitAI

AI infrastructure company Jiyuan Lüdong recently completed tens of millions of dollars in financing, and simultaneously launched a public beta of its domestic one-stop multi-model API service benchmarked against OpenRouter. A single API key can be used to call multiple models, and it has now accumulated 54,000 users, with daily Token call volume exceeding 500 billion. The company not only focuses on API aggregation, but also bets on the intelligent routing infrastructure between models and agents, and has formed a development path of multi-category product collaboration.

Industrial agents are not “shell” large models! Siemens pours century of experience into industrial AI

QbitAI

This content explicitly refutes the misunderstanding that industrial agents are just wrappers around large models, and puts forward Siemens’ core idea of deeply integrating its century-long accumulated experience in the industrial field into industrial AI. The Xcelerator platform it launched has significant essential differences from ordinary software shelves: the latter only aims to complete product sales, while the former focuses on the needs of real industrial scenarios and supports the continuous iteration and growth of products during the actual production implementation process.

Qianwen Office launches Qwen3.8-Flash for the first time, generation speed increased by 100%, Token consumption reduced by 75%

QbitAI

On August 26, Qianwen Office officially launched the Qwen3.8-Flash model and its supporting standard mode for the first time. This trillion-parameter model has been specially optimized for office scenarios, paired with agent collaborative optimization strategies, and its performance exceeds Claude Opus 4.6. The standard mode doubles the generation speed and reduces Token consumption by 75%, which can cover 95% of daily office needs, breaking the “impossible triangle” of AI performance, cost and speed.