Hugging Face Daily Papers

Maglev: Sliding Recurrent Memory

HF ★ 13 · Bo Liu, Qiang Liu · HF Mirror

This paper proposes Maglev, a recurrent Transformer architecture with fixed memory capacity that generalizes sliding window attention and supports parallel training. It adopts a dual-coupled structure: the highly expressive prefiller Q generates memory targets using hybrid attention, while decoder P only relies on sliding window attention + recurrent key-value injection for prediction, aligned with Q via memory consistency loss, and only P is called during inference. Experiments show its performance outperforms similar baselines, and parameter sharing between Q and P reduces parameter count while retaining most performance gains.

Thought-Level Beam Search for Reasoning

HF ★ 14 · Lijie Yang, Hongyin Luo, Jiawei Zhao… · HF Mirror

To address the pain point of low computing power allocation efficiency during testing of large reasoning models, where existing parallel sampling has memory bottlenecks and pruning schemes have insufficient hardware utilization, this study proposes the Gambit inference algorithm, which adopts thought-level beam search, dynamically concentrates computing power on optimal reasoning paths via a lightweight scorer, while maintaining high hardware utilization. Under the same hardware constraints, its accuracy improves by up to 6.7%, throughput exceeds 2x, token consumption reduces by up to 68.5%, outperforming all existing baselines comprehensively.

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

HF ★ 9 · Kabila Haile Soboka · HF Mirror

To address the low efficiency of manual CT rib fracture localization, the RibAssist 3D system was developed, which identifies fractures based on orthogonal dual projections generated from CT, triangulates after cross-view matching, and completes selective 3D localization under a conservative strategy. Actual measurement shows a median localization accuracy of 1.49mm, 93% of fractures are matched to the correct rib, with only 0.436 false positives per case, and the low recall rate comes from high-confidence screening. The study verified the framework is reproducible, and identified cross-view matching as the core performance bottleneck.

Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation

HF ★ 13 · Hmrishav Bandyopadhyay, Xuanchi Ren, Zijian Huang… · HF Mirror

To address the problem in autoregressive video few-step distillation where bidirectional teachers rely on future information, leading to misalignment between supervision and the student’s causal information set, the Context-Matched Distillation (CMD) framework is proposed, which uses causal teacher scoring that only uses historical information, paired with prefix matching scoring and prefix perturbation strategies to stabilize training, adapting to multiple generation scenarios. Experiments show its performance on both long and short videos reaches SOTA among autoregressive methods, and the fit of time-varying camera control is significantly improved.

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

HF ★ 10 · Yuanhe Zhang, Weiliu Wang, Jie Ren… · HF Mirror

This paper focuses on the previously unaddressed security risk of human-inaudible low-frequency signals in large audio language models (LALMs), and proposes the black-box attack method ILL, which can be implemented with general waveform templates, reducing model accuracy by up to 67 percentage points, and is almost undetectable by humans. A defense scheme DRG is also proposed, which increases the accuracy of attacked models from 28.5% to 46.1% by detecting low-frequency distribution shifts and requesting secondary recording, providing a foundation for audio robustness research.

OpenAI

The builder’s guide to GPT‑5.6

OpenAI

This GPT-5.6 Developer Guide provides implementation guidance for AI entrepreneurship scenarios: startup teams can build AI agents via two solutions: one is adopting an intelligent model selection strategy adapted to business scenarios, the other is calling the exclusive capabilities of the new Responses API added to GPT-5.6. The combination of the two methods can greatly improve development efficiency, while significantly reducing R&D and O&M costs, providing a practical path for startup teams to quickly launch cost-effective AI applications.

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

OpenAI

OpenAI has launched the new “Ultrafast Preview” API service tier, running the GPT-5.6 Sol large model with technical support from Cerebras, with inference speed up to 14x that of the original version, and peak output rate up to 750 tokens per second, which can greatly reduce the response time of large model generation, adapting to the needs of generative AI scenarios with high real-time requirements.

Anthropic News

Introducing Claude Opus 5

Anthropic

Anthropic’s newly launched Claude Opus 5 is a stepwise iteration of the Opus flagship large model. Core upgrades focus on high-value implementation scenarios: it greatly strengthens the underlying support for long-running agents, can stably carry complex multi-step task flows, and has significantly improved processing performance for coding development and professional work in various fields, providing a more reliable capability base for professional-grade AI applications.

Inviting hard questions

Anthropic

This work launches a public AI question collection initiative: it solicits the public’s most concerned and most difficult to answer questions in the artificial intelligence field from the whole society, and promises that when responding to and answering the collected questions later, it will fully disclose all work processes such as derivation and demonstration, making the process transparent and traceable, to respond to public concerns about AI development and open up communication channels between the technical side and public cognition.

Google DeepMind

Introducing Gemini 3.7 Flash

Google DeepMind

This work launches Google’s new generation of lightweight multimodal large model Gemini 3.7 Flash. The model optimizes the low-latency inference architecture, is 2x faster than the previous generation 1.5 Flash, supports a context window of up to 1 million tokens, its multimodal and programming capabilities are on par with the higher-end 3.7 Pro in the same series, and inference cost is only 1/10 of the Pro version, which can adapt to edge deployment and high-concurrency real-time interaction scenarios, with outstanding cost-effectiveness advantages.

Putting sign language AI into users’ hands

Google DeepMind

This study focuses on the implementation of sign language AI, targeting the communication needs of deaf and hard-of-hearing users, and has developed a breakthrough sign language to text (SL2T) model. This model will serve as the core technical base to support the development of new sign language functions for hearing-impaired groups, truly delivering sign language AI capabilities to users and reducing the cross-modal communication threshold for hearing-impaired groups.

Hugging Face Blog

State of Open Models: Summer 2026 Observations

Hugging Face

At present, only the title of the report State of Open Models: Summer 2026 Observations is provided, key information such as core research content, adopted methods and conclusions of the abstract are missing, so translation, refinement and a ~120-word summary cannot be completed. Please supplement the full original content of the abstract, and I will highlight the methods and conclusions as required to output a concise and clear summary.

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

Hugging Face

This article introduces the one-stop embodied intelligence development solution launched by Hugging Face, which integrates the Strands Agents development framework, LeRobot robotics tool library and official storage bucket capabilities, covering the entire process of data recording, model training, and deployment, no cross-tool switching is required, and shared datasets and model assets on the platform can be directly reused, greatly reducing the threshold for robot development and improving R&D collaboration efficiency.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This AI alignment study starts from the perspective of virtue ethics, refutes the common presupposition that “rational agents need to be anchored to fixed ultimate goals”, and points out that the rationality of human actions stems from matching the practice network containing elements such as action rules and evaluation standards, rather than pointing to fixed goals. The study proposes that to achieve compatible collaboration between AI and humans, while taking into account ethical fit and safety requirements, AI decision-making logic needs to be isomorphic with human practice-based action logic.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

The concept of recursive self-improvement (RSI) first originated from the 1965 superintelligent machine hypothesis proposed by I.J. Good: it refers to agents that can surpass all human intellectual activities, and can also independently design better systems to achieve self-iteration. In 2008, Eliezer Yudkowsky clearly defined it as the feedback loop where AI optimizes its own cognitive mechanism relying on existing intelligence. Currently in the AI field, this type of feedback includes both directly rewriting its own weights, and can also be extended to optimizing training pipelines.