Hugging Face Daily Papers

Maglev: Sliding Recurrent Memory

HF ★ 3 · Bo Liu, Qiang Liu · HF Mirror

This paper proposes Maglev, a recurrent Transformer architecture with fixed-length memory that generalizes sliding window attention and supports parallelization during training. It consists of two coupled modules: the pre-filler Q combines full attention and sliding window attention to generate memory targets, and the decoder P only uses sliding window attention plus recurrent K/V injection for next-token prediction, aligning the outputs of the two modules with memory consistency loss. Only P needs to be run during inference. Experiments show it outperforms similar baselines, and Q-P parameter sharing reduces parameter count while retaining most performance gains.

Thought-Level Beam Search for Reasoning

HF ★ 3 · Lijie Yang, Hongyin Luo, Jiawei Zhao… · HF Mirror

To address the low efficiency of computing resource allocation during the testing phase of large reasoning models, where existing parallel sampling methods are prone to memory bottlenecks and pruning methods have insufficient hardware utilization, this paper proposes the Gambit inference algorithm, which implements thought-level beam search. It dynamically concentrates computing resources on high-quality reasoning paths through a lightweight scorer, maintaining high hardware utilization. Under the same hardware constraints, its accuracy is improved by up to 6.7%, throughput is more than doubled, token consumption is reduced by up to 68.5%, comprehensively outperforming existing baselines.

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

HF ★ 0 · Kabila Haile Soboka · HF Mirror

To solve the time-consuming problem of manual localization of CT rib fractures, this study proposes the RibAssist 3D solution, which generates anterior and lateral orthogonal projections from CT scans. Fractures are independently detected from the two views and then matched, and 3D localization with controllable false positive rate is obtained via triangulation. Experiments show that the core bottleneck is cross-view matching accuracy. Under a conservative strategy, the number of false positives per case is 0.436, the median localization error is 1.49mm, and 93% of fractures can be accurately mapped to the target rib, providing a reproducible framework for related research.

Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation

HF ★ 4 · Hmrishav Bandyopadhyay, Xuanchi Ren, Zijian Huang… · HF Mirror

To address the supervision misalignment problem in existing autoregressive video distillation, which uses bidirectional teachers that rely on future information unavailable during generation, this paper proposes the Context-Matched Distillation (CMD) framework: it uses a causal teacher that only relies on historical information, paired with prefix scoring and perturbation strategies to align supervision boundaries. This method is suitable for multiple generation scenarios, achieves SOTA performance 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 focuses on the overlooked security risk of inaudible low-frequency signals to Large Audio Language Models (LALMs): it proposes ILL, a black-box red team testing method to construct covert attack signals that can reduce the task accuracy of 6 LALMs by up to 67 percentage points, and are almost undetectable by the human ear. A corresponding defense solution DRG is also proposed, which triggers resampling when low-frequency distribution shift is detected, increasing the average accuracy after attack from 28.5% to 46.1%, filling the research gap in this field.

arXiv cs.LG

LoKiFormer: Locality-aware Attention with Decoupled Knowledge Memory for Efficient Large Language Model Pretraining

Qiuwu Chen, Zimo Liu, Yuchen Li…

To address two major pain points of low pre-training efficiency of LLMs: self-attention lacks local inductive bias and tends to redundantly model local information, and the coupling of knowledge storage and computing paths in MoE architectures hinders flexible access to global knowledge, this paper proposes the LoKiFormer architecture. It adds convolution-integrated local fusion attention to capture local features, and a parameterized key-value memory module to decouple storage and computation. Experiments show that its pre-training convergence is 1.33 times faster than baselines, with better efficiency.

Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods

Matthew Kahn, Milan Arjel, Nirmala Adhikari…

To fill the gap in prediction models for three types of disasters in the Nepalese Himalayas: glacial lake outburst floods, landslides, and ice floods, this study uses only free satellite radar and meteorological data, combined with thousands of disaster records, and builds models through rigorous spatial cross-validation. The results show that the ROC of pre-disaster weather for disaster trigger timing prediction reaches 0.73-0.83, the prediction accuracy of terrain vulnerability is limited, deep learning does not outperform simple gradient boosting baselines, and a disaster monitoring priority list is finally output.

MARCH: Scaling Recurrent Memory with Content-Routed State Anchors

Ming Zhang, Kaisen Yang, Shu Yu…

To address the pain points of high long-context overhead of Transformers and easy loss of early information in fixed states of traditional recurrent models, this paper proposes the MARCH architecture: it periodically caches recurrent state anchors and associates them with content keys, and aggregates the current state and historical anchors via attention, enabling flexible trade-off between historical resolution and VRAM cost. After pre-training, it outperforms multiple linear attention variants on tasks such as commonsense reasoning and LongBench, balancing long-term memory capability and computational efficiency.

OpenAI

The builder’s guide to GPT‑5.6

OpenAI

This Builder’s Guide to GPT-5.6 focuses on the AI implementation needs of startup teams, and systematically explains the practical solution for building AI agents based on GPT-5.6: core technical paths include intelligent model selection strategy and the new Responses API, which can greatly shorten the AI agent development cycle, while significantly reducing development and operation and maintenance costs, providing a directly reusable implementation direction for startups to quickly launch cost-effective AI applications.

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

OpenAI

OpenAI has recently launched a preview version of the Ultrafast API service tier for developers, adapted to the latest GPT-5.6 Sol LLM, with underlying computing power supported by Cerebras. The inference speed of this service can be up to 14 times that of the original version, with a peak output of 750 generated tokens per second, which can greatly reduce LLM response latency, and is suitable for AI generation scenarios with high concurrency and strong real-time requirements.

Anthropic News

Introducing Claude Opus 5

Anthropic

The newly launched Claude Opus 5 is a stepwise upgrade of the Opus series LLMs, with core capabilities achieving breakthroughs in two major directions: first, the support for long-running agents is greatly improved, which can carry more complex long-process agent tasks; second, the processing performance of code generation and professional field work is simultaneously optimized, which can better meet the needs of high-level development, professional work and other scenarios.

Inviting hard questions

Anthropic

This project launches a public call for high-value difficult questions in the AI field, promising that when answering all collected questions later, it will fully disclose full-process work details including problem disassembly, technical derivation, demonstration and verification, and take the initiative to accept public supervision. It aims to narrow the information gap between AI R&D and public cognition, improve AI technology transparency, and accurately respond to the public’s practical concerns about AI.

Google DeepMind

Introducing Gemini 3.7 Flash

Google DeepMind

The Gemini 3.7 Flash released by Google this time is a lightweight flagship multimodal LLM. Its architecture adopts an optimized sparse activation mechanism, paired with low-bit quantization and inference scheduling optimization, supports a 1M token super-long context window, and can parse multimodal inputs including text, audio, video, and images. Measured inference speed is 2 times higher than the previous generation, call cost is reduced by 50%, and multimodal performance is close to Gemini 3.7 Pro, suitable for low-latency scenarios such as real-time interaction and edge deployment.

Putting sign language AI into users’ hands

Google DeepMind

This paper focuses on the communication pain points of deaf and hearing-impaired groups, and launches a self-developed breakthrough sign language to text (SL2T) model, whose core is to realize end-side deployment of sign language AI, providing technical support for new sign language functions for hearing-impaired users. This technology eliminates the previous reliance of sign language recognition on cloud computing power, can run directly on users’ personal devices, greatly lowers the threshold for use, and provides a convenient path for daily communication of hearing-impaired groups.

Hugging Face Blog

State of Open Models: Summer 2026 Observations

Hugging Face

State of Open Models: Summer 2026 Observations conducts multi-dimensional evaluations of more than 30 mainstream open source base models and fine-tuning models in terms of general capability, vertical adaptability, deployment cost, etc. It finds that the general capability of open source models has caught up with the level of leading closed-source models at the end of 2025, the cost-performance ratio of multimodal open source models is significantly better than closed-source solutions, the industrial adoption rate of small-parameter vertical open source models exceeds 60%, only code and complex reasoning capabilities still have a gap of about 15% compared with leading closed-source models, and compliance is the core obstacle to adoption.

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

Hugging Face

This is an integrated embodied intelligence R&D solution launched by Hugging Face, integrating three capabilities: Strands Agents agent framework, LeRobot robot development library, and official storage buckets. It can complete the full process of runtime data recording, embodied model training, and application deployment in a closed loop on the same platform, eliminating cross-tool adaptation costs, greatly lowering the R&D threshold for robot applications, and improving full-link development efficiency.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This AI alignment research challenges the classic orthogonality hypothesis, proposing that human rationality does not originate from the pursuit of ultimate goals, but is the result of actions adapting to a practical network composed of action tendencies, evaluation criteria, etc. The study argues that AI should not have preset fixed goals, and its decision-making logic needs to match the practical action paradigm of humans, in order to achieve ethical alignment and guarantee core security attributes at the same time.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This paper sorts out the research context of AI Recursive Self-Improvement (RSI): this direction originated from the “ultraintelligent machine” concept proposed by I.J. Good in 1965, and Eliezer Yudkowsky clarified in 2008 that its core is the feedback loop where AI iteratively optimizes its own cognitive architecture based on existing intelligence. Current RSI implementation in the AI field is divided into two types of paths: one is the model directly rewriting its own weights, the other is optimizing the training pipeline in a broad sense.

QbitAI

Qwen3.8-27B is now open source! Runs on consumer GPUs

QbitAI

Recently, Qwen officially open sourced the Qwen3.8-27B multimodal LLM, released under the Apache 2.0 license for free commercial use, and can be deployed on consumer GPUs. The model natively supports a 262K context window, which can be extrapolated to 1M tokens, its programming and office performance exceeds Qwen3.7-Plus, and it also adds a resource-saving function that adjusts thinking depth according to task difficulty. So far, Qwen has open sourced more than 460 models, with global downloads exceeding 3 billion times.

$100 million annual salary couldn’t keep him! Jiahui Yu leaves Meta to start a business

QbitAI

Core researcher Jiahui Yu, who was poached by Meta from OpenAI with a $100 million annual salary, has only been in the position for more than a year. After leading the core multimodal team of the Superintelligence Lab to complete the R&D of the full-stack Muse series products, he recently officially announced his departure to start a business. He revealed that his new direction is an unexplored field of great significance to the future of humanity, details of the new company are not yet public, and will be disclosed after the project takes shape.

Google begins to dismantle DeepMind, several teams are transferred to headquarters

QbitAI

After DeepMind’s leadership change, Google launched the split. At the all-hands meeting on August 6, it announced that several of its non-technical support teams would be transferred to the headquarters reporting system, while research teams remain, aiming to reduce intermediate links and speed up the integration of Gemini into products such as Search and Cloud. Google co-founder Sergey Brin has returned to intervene in Gemini R&D, promoting resource allocation towards cutting-edge models and recursive self-improvement. Details of the adjustment are not yet public, and employees have doubts about subsequent project resource allocation.