Baseten Labs6565Claude Code v2.1.159 released, adopting a gradient alignment reward mechanism, using the Qwen3-8B model on arithmetic tasks, rewarding by calculating the gradient similarity between student output and expert demonstrations, experiments show the model may overfit the precise wording of expert examples.
NVIDIA Technical BlogTwo NVIDIA DGX Stations connected via ConnectX-8 RDMA, using vLLM and tensor parallel processing, successfully ran a 175B parameter GLM-5.3 NVFP4 model.
Ai2 Blog4040XYBench released, including 8K+ problem tests to evaluate how AI assistants handle requests based on misunderstandings; Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, and DeepSeek-V4 Pro often focus on the original incorrect method rather than the user's actual goal in the test.
Hugging Face Daily PapersOpen Vision Banana (OVB) has open-sourced its data, weights, code, and implementation details. Its 4B and 9B parameter-based open-weighted large model and 2.1M-3.8M training pairs surpass Vision Banana on Cityscapes with zero-shot performance of 73.9 vs 69.9 mIoU and can rival state-of-the-art affine-invariant depth estimators.
Hugging Face Daily Papers6565Decision models, even as actors, can effectively teach reinforcement learning strategies by evaluating complete action trajectories, improving success rates on ALFWorld and WebShop.
Hugging Face Daily PapersDuo v1 introduces the Reciprocal World Model (RWM), which jointly models intentions and reality through autoregressive reasoning, synchronized attention, and dual inference distillation, achieving real-time full-duplex audio-video interaction. The model continuously perceives audio-video input while responding to users in real time, treating intelligence and physical reality as equal foundational components of the world, supporting interactions such as dialogue, sharing objects, and reshaping virtual scenes. The first batch of models has been open-sourced.
Hugging Face Daily PapersBirdDreams jointly generates bird's-eye view (BEV) videos and bird's-eye view videos, making scene layout a part of predicting the future. By providing an overhead view of vehicle position and movement through BEV, it reduces occlusion, while the perspective video retains scene details. On PhysicalAI-AV and nuScenes, BirdDreams more faithfully follows the input trajectory of its own vehicle and generates more reasonable self-motion without HD map or vehicle state input.
OpenRouter BlogWhen RAG applications exist correct information in documents but provide wrong answers, it is usually due to errors in the retrieval stage. Vector search can find relevant passages on the topic, but it often fails due to approximate matching, such as incorrectly calling the bucket limit of the Enterprise plan when asking about the Growth plan.
Google Developers BlogGemini 4 Argon translates Go code panic paths into Z3 constraints, paired with Z3 to check trigger conditions, and paired with go test to replay trigger bytes, locating 3 embedded parser bugs and discovering 3 skill-lint potential bugs within 1 minute. This process is more efficient than go test -fuzz, capable of finding boundary check errors missed by fuzz testing.
PyTorch Blog5050Helion analyzes kernel IR and data flow through compile-time heuristic algorithms, selecting fast settings without slow automatic tuning, classifying kernels as matrix multiplication, reduction, or point-wise operations, and choosing block size, warps, and phases to achieve near-automatic tuning speed.
AWS Machine Learning Blog5858Amazon Bedrock Managed Knowledge Base achieves fine-grained access control for enterprise AI agents by injecting user context into the AgentCore Gateway tool layer and verifying JWT, combined with the ACL-aware functionality of the knowledge base.
IBM Research BlogIBM Research is studying two quantum computing hardware architectures, one of which is the spin qubit architecture that encodes information by controlling the spin state of electrons. Spin qubits use quantum dot structures (such as silicon-germanium heterostructures) to trap individual electrons, controlling their spin direction (up or down) with voltage pulses to represent 0 or 1. This architecture has high gate fidelity, scalability, relatively fast speeds, and simple control signals (requiring only voltage pulses), but the current number of qubits is less than that of superconducting qubits and faces challenges such as uniformity and manufacturing yield.
Galaxy Smart Learning has launched an AI education product, Xiao Lu Ai Xue, which adopts a dual-core architecture of the Doubao large model and a self-researched intelligent learning engine. It achieves learning situation diagnosis by analyzing students' answer records and develops functions such as learning situation diagnosis, ability assessment, learning difficulty matching, and task planning. The student-side intelligent agent recommends learning tasks and explains the basis, while the teacher-side intelligent agent summarizes student performance for personalized guidance reference. The product has served over 500 schools, covering 24 provinces, and supports teachers in adjusting teaching strategies through a human-machine co-education model.
PyTorch BlogThe EUROPA Union successfully trained a MoE model with 951.9B parameters on the JUPITER supercomputer using PyTorch and TorchTitan, increasing single GPU throughput from 90 to approximately 1,300 tokens/s with 384 NVIDIA GH200 GPUs.
OpenRouter BlogRunning the same tasks on OpenRouter, GPT-6.1 Sol passed 180/180 tests in 60 synthetic tasks, DeepSeek had the lowest cost passing 157/180 times, but the correlation with the public ranking is low (Spearman correlation coefficient 0.109). Need to check the number of task repetitions, the number of answers, and billing verification.
Google Research5858Google Research releases a wearable device-driven CKM risk assessment framework, predicting future health trajectories through an explainable health age model (CHA). The framework is trained on wearable data from 2.2 million weeks using the SensorWM model, providing personalized 12-week counterfactual health simulations combined with language models to generate clinically validated guidance plans. The CHA model integrates modifiable indicators such as heart rate, steps, and sleep to predict disease risk over 68,000 person-years, improving the predictive performance of clinical indicators.
Apple Machine Learning ResearchState Space Models (SSMs) lag behind Transformers in contextual learning and precise retrieval tasks, as their gating mechanism leads them to first learn memory solutions, delaying or preventing the convergence of correct contextual learning solutions, but the gating mechanism is beneficial for improving long sequence generalization.
Hugging Face Blog3838Hugging Face's blog introduces the Token Pooling perception training method for multi-vector embedding models, which compresses storage by merging similar tokens through HierarchicalTokenPooling.
vLLM Blog5858vLLM extends multi-modal service capabilities through the EPD (Encoder/Prefill/Decode) separation technology, supporting independently deployed visual encoders. vLLM v0.32.0 includes MiMo and XPU support, and the metadata-only request feature of DeepSeek V4.1 Flash has been open-sourced.
Ant Group inclusionAI · Hugging Face new modelsinclusionAI has released the ZoomGain-4B large language model, achieving a score of 61.08 on ZoomGainBench and performing excellently on multiple fine-grained perception and general perception benchmark tests; the model is built with supervision based on the global-to-local visual perception gap mined from global image details, supporting English and Chinese Q&A, and can be downloaded via Hugging Face.
Hugging Face Daily Papers6565SWAM extends the World Action Model through cross-modal alignment, trained on 47,000 hours of multimodal data, blending egocentric video and UMI demonstrations to enhance control generalization, ultimately achieving a 40.00% success rate on three real-world tasks, surpassing the baseline of 23.33%.
Hugging Face Daily Papers6565The DexForesight framework implicitly learns spatial foresight by injecting a dynamic perception world representation into the latent space of dexterous strategies, increasing the success rate of basic strategies from 49.4% to 60.5% on 11 DexJoCo simulation tasks and achieving a 58.3% success rate on 3 real robotic dexterous operation tasks.
Hugging Face Daily Papers6565Infrathon conducted 150 long-term AI infrastructure optimization task evaluations on six models, consuming a total of 10,000 GPU hours, and found that the model rankings changed with the evaluation duration. 135 tasks achieved further optimization in the latter half of the budget, and Claude Opus 5.5 increased the KV cache compression rate from 6.82% to 19.86%, exceeding human design baselines. Analysis shows that task violations, unreliable testing, and premature completion are the main obstacles to continuous optimization.
Hugging Face Daily PapersdNaViT is a discrete native resolution vision Transformer that matches or exceeds specific baselines in visual understanding, generation, and editing tasks through grouped product quantization and dual-branch targets. It performs excellently on 12 visual understanding benchmarks, 4 generation benchmarks, and one image editing benchmark, proving that discrete vocabulary can be a universal visual representation
OpenRouter BlogThe GitHub MCP server added 19,436 input tokens for each request of Claude Sonnet 5.5, far exceeding the CLI (1,421) and skills (3,315).
OpenRouter BlogTesting on OpenRouter, fp8, fp4, and fp16 showed no precision differences in 240 scoring tasks, with all six endpoint scores ranging from 98.3% to 99.2%. Lower precision is generally cheaper, but there were 7 models where fp8 was more expensive than fp16, and 7 models where 4-bit was more expensive than 8-bit. The fp4 endpoint performed the worst among 40 workflows without multiplication tasks, but the difference might be due to chance.
AWS Machine Learning Blog4040Amazon Bedrock AgentCore Evaluations provides three evaluators—built-in, DeepEval, and AutoEval—to assess agents. The built-in evaluator is developed and tested by AWS, covering common dimensions; DeepEval and AutoEval are third-party open-source libraries, with AgentCore hosting them in production environments. The evaluation process requires first defining business success criteria, breaking them down into observable behaviors, and then mapping them to the appropriate evaluator level.
OpenAI Alignment BlogGPT-6 Astra shows a decrease in CoT monitoring ability in non-adversarial settings, an increase in CoT evasion ability in adversarial tests, and no significant improvement in CoT capability. During pre-training, CoT capability does not increase with the decrease of the base model's perplexity, and the architecture has a limited impact. Reinforcement learning enhances CoT control ability, and model-generated data fine-tuning can transfer CoT control ability. Model perplexity is positively correlated with CoT capability and control ability.
The DecoderNvidia's GR00T 2 AI model learns new tasks for robots by analyzing human video actions without remote control, trained on 400,000 hours of video data, with lab test success rates of 10-80%, and combined with GPT-6 Astra achieves 97% success rate for toy model assembly, but tactile processing and language model latency remain challenges.
Meta EngineeringMeta open-sources Triton Low-Level Extensions (TLX) compiler, enhancing GPU performance by introducing explicit warp-group and cluster-level orchestration, supports RecSys training/inference and LLM pre-training, achieving end-to-end performance improvements over standalone libraries on NVIDIA and AMD GPUs.
Hugging Face Daily PapersFrontier3DBench uniformly evaluated the ability of code agents in object, scene, and interactive 3D world creation across 145 tasks, GPT-6.1-sol performed best at the object and scene levels, with a total score of 66.02/100, but all configurations had reference reconstruction scores lower than text-driven generation in object and scene tasks. The generated world averaged 83.08/100 in local mechanisms, but only 20.82 and 30.26 respectively in motion removal and stateful workflow, showing that coordinated activity completion remains a core bottleneck
openJiuwen releases the TTSE (Dual-Track Self-Evolution) framework, dividing Agent experience into FACT (Environmental Facts) and TIP (Conditional Techniques) managed on dual tracks, continuously calibrating environmental cognition and execution methods through real task feedback.
Hugging Face Daily PapersLingBot-Touch is a tactile foundational model that learns a unified, scalable representation across different tactile sensors (resistive, magnetic, visual tactile), producing plug-and-play tactile tokens, and continuously enhancing the dexterity of diffusion strategies and visual language action (VLA) models (such as π_{0.5} and LingBot-VLA). Experiments show that this tactile representation significantly improves performance in complex manipulation tasks in both simulation and real robot deployment while maintaining strong computational efficiency.
Hugging Face Daily PapersA 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.
Hugging Face Daily PapersOneTrans-GSR generates unified recommendation ranking lists through a single autoregressive Transformer, improving business metrics and reducing re-ranking and hybrid sorting latency. After deployment in industrial recommendation systems, GMV/u increased by 2.63%, advertiser value increased by 2.06%, and repurchase rate increased by 2.03%, while end-to-end latency was reduced by 13%.
Hugging Face Daily PapersWorldFoundry provides open-source infrastructure for reproducibly reasoning and evaluating generative world models across video, 3D/4D, embodied action, and interactive systems, including traceable evidence and benchmarks, featuring 270 model entries, 74 benchmarks, and 62 registered metrics.
Anthropic ResearchAnthropic partnered with Diode Computers to test the ability of Claude Fable 5.2 to design electronic devices, a model capable of completing electronic hardware from design to prototyping.
Hugging Face Daily PapersPCBRouteAgent 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.
