#Hugging Face

Fast and Exact Relative-Position Attention via Structure-Aware Kernels

A precise tile attention algorithm integrates additive relative position terms into the computation similar to FlashAttention, reducing memory and computational overhead, accelerating long-context relative position Transformers. The method (1) uses offset-aware tiles to avoid redundant index and memory access, (2) only computes necessary key/query relation products and reuses them in forward and backward passes, and (3) uses offset-based gradient sharding and aggregation to eliminate atomic contention.

LLMs Don't Know What They Know: Benchmarking Temporal Self-Knowledge for Financial Backtesting

GPT-5 correctly answered 76% of control facts in financial backtesting but claimed to know only 38%, and still correctly answered 61% of the facts it claimed to be unaware of. The claimed knowledge can extend to seven quarters after the measurement knowledge boundary. Historical date instructions can eliminate most knowledge claims without changing accuracy.

PCBRouteAgent: How Far Are LLM Intelligent Agents from Becoming PCB Routing Engineers?

PCBRouteAgent achieves 100%, 94.9%, 100%, and 40% D2, D3-A, D3-B, D3-C cleanliness pass rates on 247 synthetic and real PCBs, being 13.9, 36, and 10 percentage points higher than the best baseline on three real-world subsets respectively.

LowRAM: A runtime method for low-rank keys and amortized values for composable KV cache compression

LowRAM compresses keys using low-rank decomposition and uses small head networks to predict key values, while retaining residual correction for the most difficult-to-reconstruct markers, achieving 4x value cache compression with near-lossless performance. Selective reconstruction can increase peak decoding throughput by 5x for 262K context under 4x KV compression. Combining quantization and eviction, LowRAM pushes compression to 26.9x while retaining 94.5% of baseline performance.

LongBanana: The first expert-verified long-context multi-reference image synthesis benchmark

LongBanana is the first expert-designed and validated long-context multi-reference image synthesis benchmark, containing 90 carefully designed tasks, each task containing about 32 reference images, 20 constraints, and 50 evaluation items, a total of 4,670 expert-specified atomic evaluation items.

Drex open-source decision model released

Drex open-source decision model returns the probability distribution of user options without autoregressive decoding, using supervised fine-tuning, RLAF, and consistency loss to improve accuracy, robustness, and latency. Its 35B parameter Max version achieves an industry-leading score of 67.52 on Decision Index 0.4.1, with a response time only 9% of GPT-6 Sol and 5% of Claude Opus 5.5.

SLA: Expanding Spatial-Language-Action Model via Continuous Pre-training

SLA model aligns pre-trained VLA strategy with frozen 3D perception model spatially through continuous pre-training and post-training, using feature-dependent gating and probabilistic state dropout to enhance spatial understanding across different robot datasets while retaining the inference architecture.

Hugging Face Open ASR Leaderboard adds Egyptian and Nigerian English datasets

Hugging Face Open ASR leaderboard adds Liva AI private datasets for Egyptian English and Nigerian English, containing 1.34 hours and 1.55 hours of dialogue speech, to enhance the representativeness and reliability of ASR evaluation. The leaderboard now supports more detailed accent classification based on the Kachru three-circle model, covering groups such as US and non-US, inner circle and outer circle/extended circle.

MUON-RT optimizer reuses teacher trajectories to implement momentum-free LLM self-distillation

The MUON-RT optimizer distills teacher model information to the student model online by reusing teacher trajectories and removing momentum weights, saving 25% memory parameters (standard Muon requires 4x parameters, MUON-RT only requires 3x), improving model performance.

WeLM-Image: A unified image generation model for visual intelligence

WeLM-Image is a unified image base model that integrates multimodal understanding, supporting text-to-image generation, instruction-guided editing, and multi-reference synthesis.

Harness-S: Learning from sibling Harness to achieve general intelligent agent training

Harness-S achieves general intelligent agent training by learning from its sibling Harness. The model scores 82.5 on the MMLU benchmark, 3.2 times faster than baseline models, with a context length of 128k tokens and a parameter size of 175B. The model is open-source and can be accessed via API calls.

Hugging Face CEO Delong: AI 'doomsday risk' is close to zero, closed development is the real danger.

Phonon TTS releases on-device speech models, topping multiple metrics on the Hugging Face open TTS leaderboard

Phonon releases a small TTS model that can run on local devices, without a server and network, offering two versions of about 65 MB and 75 MB, covering English, French, German, Spanish, and Portuguese.

The NEO-VLA framework eliminates fixed visual encoders to enhance the robustness of action strategies

NEO-VLA framework eliminates fixed visual encoders by allowing action strategies to directly access deep visual representations in the network, achieving stronger robustness in various manipulation scenarios and distribution shifts. The framework uses lightweight convolutional layers to tokenize images and allows visual representations to evolve together with language in the backbone network, enabling action experts to obtain information from different depths of instruction-conditioned visual representations. Layer exploration experiments show that action key signals often reach their peak before the final visual representation is formed.

How does multi-round reinforcement learning improve LLM agent performance?

Multi-round reinforcement learning (RL) performs excellently on long-term tasks, but performance gains highly depend on interaction constraints. By evaluating different sampling widths K and interaction depths T using Pass@(K,T), it was found that the advantage of trajectory-level RL depends on sufficient episode budgets, and the advantage significantly decreases when the budget is tightened.

The TurnWise framework enhances streaming video understanding through round-level credit assignment.

The TurnWise framework improves streaming video understanding by round-level credit assignment and decision weight balancing. Its round-level credit assignment constructs group relative advantages using prefix scores from silent completions, and round-mean policy optimizes the token loss per round. Experiments show this method outperforms existing streaming video understanding RL methods.

D3D: Decoupled Diffusion Drafting Asynchronous Speculative Decoding Technology

D3D breaks the dependency on the hidden states of the target model through an independent diffuser, achieving asynchronous drafting and verification. In seven benchmarks, D3D combined with block denoising and auxiliary autoregressive supervision increases the average decoding speed of Qwen3-8B and Qwen3-32B by 5.00 times and 5.71 times respectively, outperforming DFlash (4.03 times and 3.67 times), and reaching 7.48 times acceleration on GSM8K. The code, model checkpoints, and benchmark scripts are open source.

webAI has released TwIL-8B, an 8.8B inference model that scored 98.2 on MATH-500.

webAI has released TwIL-8B, an 8.79B dense inference model fine-tuned based on IBM Granite 4.2 8B, achieving 98.2 on MATH-500 and 81.6 on LiveCodeBench†, using 38% fewer tokens than Qwen3.5-9B.

VidUEU-Agent: An intelligent agent for data management in multimodal understanding, editing, and unification tasks.

The VidUEU-Agent framework solves the problem of scarce high-quality multimodal instruction data and high annotation costs in multimodal model development by automatically constructing instruction data from raw videos. The framework adopts a four-stage data construction process, combining visual language models and visual tools for evidence-based data synthesis, and utilizes downstream training feedback for autonomous iteration.

Route Anything: A Research Framework for Model and Strategy Routing in General AI Systems

The Route Anything framework enhances the performance and efficiency of general AI systems by dynamically selecting and coordinating different AI models and strategies. It has constructed a 250,000-run trajectory matrix covering 25 benchmark tests, including 452 configurations such as language understanding, image, video, speech generation, coding, task proxying, and robotic operation.

YANAMI: A Large-Scale Multimodal Dataset and Benchmark for Multi-Round State Interaction

YANAMI introduces a large-scale multimodal dialogue dataset and benchmark that requires models to track and implement evolving dialogue states across rounds, revealing that current unified multimodal models still struggle with this task. YANAMI-400K contains 400,000 multimodal dialogue rounds across 4 scenarios, requiring models to continuously track, store, and revise evolving multimodal dialogue states across rounds; YANAMI-Bench contains 400 multimodal rounds for system evaluation of multimodal state interaction, jointly evaluating dialogue state understanding and visual implementation.

The same model with a different harness: Comparing CUGA, Hermes, and Deep Agents on AppWorld.

The IBM team compared CUGA, Hermes, and Deep Agents using the same open-source weight model GLM-5.3-flash on 90 tasks in AppWorld, with pass rates of 81.1%, 46.7%, and 30.0%, respectively, and 20, 8, and 2 scenarios where all three variants passed.