Articles from HuggingFace
Hugging Face is switching its Inference Providers from a custom router to OpenRouter, letting any open-weight model on the Hub be called via a single Hugging Face token with automatic provider selection based on price, speed, and reliability.
tiny-voca 是一个仅 4.5M 参数的轻量英文文本转语音模型,包含两个美式英语音色,在四线程 CPU 上运行速度约为实时的 70 倍,提供 PyTorch 与 ONNX 权重、代码及浏览器试用空间。
A walkthrough showing how to build a robot pick-and-place simulation in Isaac Sim using NVIDIA's SimReady OpenUSD assets, including the UR17G arm, Robotiq 2F-140 gripper, and a beverage can asset from the SimReady Central library on Hugging Face.
AllenAI and collaborators release XYBench, an 8K+ question benchmark (Best Resource Paper at EMNLP 2026) that tests whether AI assistants recognize when a user's request is based on the XY problem misconception and redirect toward the actual goal. Evaluations of Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, and DeepSeek-V4 Pro show models still focus on the mistaken approach rather than steering users to the underlying goal.
llama.cpp releases v1.0.0 of its libraries, establishing a stable API/ABI foundation with strict semantic versioning for the local AI ecosystem, plus expanded distribution channels and CI infrastructure.
DySENSE is a joint multimodal GenAI framework that synthesizes static urban satellite imagery alongside spatially aligned, dynamic monthly building energy consumption maps, decoupling long-term urban morphology from short-term energy dynamics. Evaluated on four cities, it achieves strong downstream performance with less than 25% labeled data.
OneSteer is a one-step diffusion whole-body controller that combines general humanoid motion tracking with selective test-time steering via differentiable task costs and reference masks, achieving a 90% clean passage rate under a 1.25m bar (up from 2%) and completing 27/27 nominal tracking trials on a physical Unitree G1 without falling.
该论文评测了73种无参考图像质量评估指标与人类偏好的一致性,发现没有指标能稳定匹配人类判断,并提出了包含人类参考、编辑、对抗攻击与生成式漂移图像的基准TrustUIQA。
V-JEPA-LM introduces open video-language models combining a frozen V-JEPA 2.1 world-model encoder with Qwen3/Llama backbones, trained in one stage with counterfactual supervision that forces models to reason about dynamics rather than static text-aligned features, leading all size classes on 11 video benchmarks with far less data and fewer visual tokens than Meta's V-JEPA 2/2.1 video LLMs.
This paper introduces 'agentic robotics' as a new paradigm where a foundation model paired with an embodied harness forms an 'Embodied System 2' for open-world physical tasks, and proposes an evaluation framework plus a research roadmap toward sustained autonomy.
研究分析了 ACL、EMNLP、COLM 和 ICLR 的 2000 篇论文,发现 GPT 和 Claude 等大模型在审稿中约一半以上的批评点来自附录,而人类审稿人仅有 8.4%,且更关注论文呈现与核心思想。
The paper introduces HARE, a history-aware Text-to-SQL benchmark, and MEM-SQL, a memory-powered method that reuses verified SQL templates and blocks with on-demand schema linking, raising mean execution accuracy from 37.1% to 50.0% across six language models while cutting input tokens by 76.6% and latency by 68.8%.
CANTraj releases a naturalistic driving dataset and benchmark for recovering vehicle trajectories from CAN signals without GPS, showing that combining steering signals with road-network constraints raises five-minute route recovery from 56% to 87%, and that an LLM harness further improves candidate route selection.
miHoYo's Honkai AI R&D team introduces Gamer, a generalist Vision-Language-Action foundation model that plays long-horizon 3D games in real time via screenshots, keyboard, and mouse—using 10K+ hours of gameplay data, a Streaming-vLLM inference engine with 20× lower latency, and Temporal-Aware GRPO for hour-long RL training, handling over 95% of Honkai: Star Rail's main story challenges.
The paper introduces JVT (Just Video Transformer), a simple large-patch DiT-style Transformer that enables pixel-space video diffusion without pretrained autoencoders or U-Net skip connections, achieving quality competitive with latent diffusion baselines and supporting joint camera-conditioned RGB-disparity generation.
eDIAR proposes a training-free, open-speaker streaming diarization method that removes fixed speaker-slot limits by clustering speaker embeddings with the e-means algorithm, achieving better accuracy on long meetings than Streaming Sortformer with constant per-update cost.
JFinBench introduces a continuously refreshable Japanese financial benchmark of 321 examples drawn from 2026 certification exams, built via a human-in-the-loop agentic pipeline to mitigate data contamination. Evaluating 12 LLMs, Claude Fable5 and GPT-6 Astra exceed 93%, while the best open-weight model Kimi K3.1 Pro trails by 12 points, with tool access narrowing but not closing the gap.
MMFlash improves multimodal speculative decoding by adaptively upweighting visually dependent tokens and using sparse attention for long video inputs, achieving up to 3.58× lossless decoding speedup and outperforming EAGLE-3 and DFlash.
IBM releases LLMON, an LLM-native markup language, along with two experimental Granite 4.1 3B and 8B preview models trained to consume and generate LLMON-marked text, improving structure handling and nearly eliminating indirect prompt injection attacks.
Haru-4B is a low-latency, multilingual full-duplex continuous spoken dialogue foundation model that listens and speaks simultaneously at 7.5 Hz, using a Qwen3.5-4B backbone with a MeanFlow audio head for one-step continuous audio generation. It is released as a research base model with PyTorch, MLX, and vLLM-Omni implementations.