Hugging Face Daily Papers

Maglev: Sliding Recurrent Memory

HF ★ 5 · Bo Liu, Qiang Liu · HF Mirror

This paper proposes the Maglev sliding recurrent memory architecture, an improved sliding window attention recurrent Transformer that supports training parallelism. It couples two modules: prefiller Q, which has access to full history and uses interleaved full/sliding window attention to generate memory targets, and decoder P, which only uses sliding window + recurrent key-value injection for inference. The two are aligned via memory consistency loss, and parameter sharing can also be adopted to reduce costs. It outperforms similar baselines on both validation loss and downstream pretraining benchmarks.

Thought-Level Beam Search for Reasoning

HF ★ 6 · Lijie Yang, Hongyin Luo, Jiawei Zhao… · HF Mirror

Aiming at the inefficient computing power allocation of large reasoning models during the testing phase, as well as flaws in existing parallel sampling and pruning solutions such as memory bottlenecks and low hardware utilization, this paper proposes the Gambit inference algorithm. It adopts thought-level beam search, dynamically concentrates computing power on high-potential reasoning paths through a lightweight scorer, while maintaining high hardware utilization. Experiments show its performance is comprehensively better than baselines: with the same hardware, accuracy is improved by up to 6.7%, throughput is more than doubled, and token consumption is reduced by up to 68.5%.

RibAssist 3D: Biplanar Rib-Fracture Detection, Addressing, and Selective 3D Localization from CT-Derived Projections

HF ★ 0 · Kabila Haile Soboka · HF Mirror

This paper develops the RibAssist 3D rib fracture localization framework: based on anteroposterior and lateral orthogonal projections generated from CT, it first detects fractures separately, then performs cross-view registration and triangulation to achieve selective 3D localization. Experiments show the core bottleneck of the system is not geometric accuracy or detection capability, but cross-view matching confidence. After optimizing the lateral detector, under conservative strategy there are only 0.436 false positives per case, the median localization error is 1.49mm, and 93% of fractures can be accurately assigned to the corresponding rib.

Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation

HF ★ 5 · Hmrishav Bandyopadhyay, Xuanchi Ren, Zijian Huang… · HF Mirror

Aiming at the problem in autoregressive video few-step distillation where the bidirectional teacher uses future information unavailable during generation for supervision, leading to misalignment with the student’s causal information set, the Context-Matched Distillation (CMD) framework is proposed. It uses a causal teacher, paired with prefix scoring and prefix perturbation strategies to align the supervision boundary. This framework adapts to multiple generation scenarios, achieves SOTA performance among autoregressive methods on both long and short video benchmarks, and significantly improves the fit of time-varying camera control.

From Inaudible Inputs to Model Failures: Low-Frequency Safety Risks in LALMs

HF ★ 2 · Yuanhe Zhang, Weiliu Wang, Jie Ren… · HF Mirror

This paper studies the previously overlooked security risks of inaudible low frequencies for large audio language models (LALMs). It proposes the black-box testing method ILL, which uses a general waveform template to construct almost imperceptible low-frequency attacks, reducing the task accuracy of 6 LALMs by up to 67 percentage points. The supporting defense solution DRG requests re-recording by detecting low-frequency distribution shifts, increasing post-attack accuracy from 28.5% to 46.1%, providing support for audio model robustness research.

OpenAI

The builder’s guide to GPT‑5.6

OpenAI

This Builder’s Guide to GPT-5.6 targets AI development scenarios for startups, and mainly introduces two practical implementation solutions for GPT-5.6: first, intelligent model selection strategies, and second, the function calling method of the new Responses API. Applying these methods can greatly improve the efficiency of startup teams building AI agents, reduce development and operation costs, and provide reusable practical guidelines for small and medium teams to implement AI businesses in a lightweight way.

Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed

OpenAI

This content introduces the new “Ultrafast Preview” API service tier launched by OpenAI for commercial scenarios. Supported by computing power provided by Cerebras, the inference speed of the GPT-5.6 Sol model can reach up to 14 times the original level, with a peak output token rate of 750 per second, greatly reducing LLM response latency, and adapting to scenarios with high real-time requirements such as generative interaction and streaming output.

Anthropic News

Introducing Claude Opus 5

Anthropic

The newly released Claude Opus 5 is a stepwise upgrade product of the high-end Opus large model line of the Claude series. This version strengthens two core capabilities: first, it greatly improves support for long-running agents, and can stably undertake complex agent tasks with long-term continuous interaction; second, the performance of code generation and professional task processing in various fields is simultaneously optimized, which can better meet the high-level work needs of developers and professional practitioners.

Inviting hard questions

Anthropic

This announcement titled Inviting Hard Questions proposes to publicly collect all kinds of difficult questions from the public about the artificial intelligence field. The team explicitly promises that in the entire process of responding to and researching all collected related questions in the future, it will fully disclose work progress, research ideas and derivation processes, take the initiative to achieve research transparency, and protect the public’s right to know about AI-related topics.

Google DeepMind

Introducing Gemini 3.7 Flash

Google DeepMind

This article introduces the lightweight multimodal large model Gemini 3.7 Flash launched by Google DeepMind. The model optimizes the inference scheduling architecture, achieves full-scenario adaptation across end devices, edge nodes and cloud via mixed-modal pretraining, LLM distillation and pruning. Its inference speed is 8 times higher than Gemini 1.5 Pro, with costs reduced by 90%. Its multimodal understanding and code generation capabilities match those of the previous generation mid-range model, supports million-level context window, and can be deployed on mobile devices to run real-time interactive tasks.

Putting sign language AI into users’ hands

Google DeepMind

This research targets the communication needs of hearing-impaired groups, and launches a breakthrough sign language to text (SL2T) AI model that can support the implementation of various practical sign language functions. This technology breaks the previous limitation that sign language AI mostly stays in the laboratory with poor implementability, directly opens sign language interaction capabilities to end users, effectively lowers the communication barrier between hearing-impaired groups and hearing people, and promotes the inclusive implementation of sign language AI.

Hugging Face Blog

State of Open Models: Summer 2026 Observations

Hugging Face

Currently, only the title of this 2026 Summer State of Open Models observation report is provided, with the full abstract content completely missing, so corresponding translation, extraction and summarization work cannot be completed. Please supplement specific content such as core methods, experimental conclusions, key findings and other details from the abstract, and I will organize it into a concise summary of around 120 words highlighting core information as required.

Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets

Hugging Face

This is a one-stop robot AI development solution launched by Hugging Face, which connects three major components: Strands Agents agent scheduling, LeRobot robot algorithm library, and official storage buckets. Relying on unified storage to host full-link data assets, it enables the entire process of data recording, model training, and deployment to be completed on the same platform, eliminating the need for cross-platform data migration and interface adaptation, greatly lowering the development threshold for robot agents and improving implementation efficiency.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This paper studies AI alignment from the perspective of virtue ethics, refutes the common presupposition that “rational agents need to be anchored to ultimate goals”, and points out that human rationality does not point to fixed goals, but adapts actions to a practical network covering behavioral tendencies, evaluation standards, and supporting resources. It proposes that to achieve AI-human collaboration and meet alignment and even safety requirements, AI’s decision-making logic must match the practical reasoning framework of humans.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This paper sorts out concepts around AI recursive self-improvement (RSI): I.J. Good first proposed the relevant concept in 1965, defining a system that can surpass all human intellectual activities and independently design better agents as an ultra-intelligent machine; in 2008, Eliezer Yudkowsky clarified that RSI is a feedback loop where AI iteratively optimizes its own cognitive mechanism relying on its existing intelligence. In the context of modern AI, RSI includes both the scenario where the model directly rewrites its own weights, and the broader scenario of optimizing its own training process.

QbitAI

Zhizhi Research Institute Proposes New Route for LLM Interpretability: Disassemble Weights, Data Cost Less Than 1%

QbitAI

Existing research on LLM mechanism interpretability mostly relies on additionally trained surrogate modules, which is costly, prone to introducing errors, and requires separate adaptation for different models. Zhizhi Research Institute, together with multiple institutions, proposed a new method called Sparse Weight Decomposition (SWD), which directly disassembles intervenable units from pretrained weights without the need for additional training of surrogate networks. The data cost is less than 1% of similar training-based methods, and it has higher efficiency in task circuit extraction for multiple mainstream LLMs.

Curing Common AI Music Flaws! This Domestic Music Model Directly Challenges SUNO

QbitAI

Currently, AI music generally has the common problem of being unable to perceive users’ delicate emotions and fragmented inspirations, with generated content being templated and poorly controllable, even the international benchmark SUNO has not broken through this bottleneck. Yinchao, a domestic full-stack self-developed team that has produced the official theme song for WAIC for two consecutive years, launched the fundamentally reconstructed V4.0 music large model, which was launched on all platforms on August 14. It has solved the problem of emotional expression recognition, directly challenges SUNO, and breaks the long-term follow-up situation of domestic AI music.

Yuan Shen Launched! Run “Opus-level” Agents on One Consumer-grade GPU, Qwen 3.8-27B Outperforms Claude on Multiple Leaderboards

QbitAI

Tongyi Qianwen’s latest open-source Qwen 3.8-27B large model has 27 billion parameters, natively supports multimodality and 262K context window, with a focus on strengthening coding and long-range agent capabilities. It significantly outperforms Claude Opus 4.6 Max on software engineering and agent evaluation leaderboards such as SWE-bench Pro. After quantization, it can run on a consumer-grade graphics card with 24GB of VRAM, enabling local deployment of Opus-level agents, and has been widely popular among developers since its release.