New releases, benchmarks, and capability breakthroughs - every model that matters.
This paper introduces Selective Sinkhorn Routing (SSR), a lightweight method for assigning tokens in Sparse Mixture-of-Experts (SMoE) models. SSR optimizes expert utilization through optimal transport constraints, significantly improving model performance and training efficiency.
Watch 9 videos detailing the real-world capabilities of the Gemini Omni and Gemini 3.5 models announced at Google I/O 2026.
Anthropic has launched Claude Fable 5, its first Mythos-class model publicly accessible, equipped with safety guardrails to prevent high-risk outputs in sensitive areas like cybersecurity and biology.
Engineers at Nextdoor are utilizing advanced AI models like Codex and GPT-5.5 to effectively tackle complex investigation issues, cross-platform building, and product outcome focus.
This research investigates how augmenting small fMRI datasets with synthetic data from a large, pre-trained model (TRIBE v2) can significantly improve the accuracy of image decoding, showing up to a 68% gain over real-data-only training.
This research introduces a novel framework using Bayesian Neural Ordinary Differential Equations (ODEs) and Gaussian Process (GP) priors to accurately predict vessel trajectories and quantify uncertainty from irregular AIS data, addressing the challenges of missing reports and co
This paper introduces NF-CoT, a framework that utilizes Normalizing Flows (NF) to model intermediate reasoning steps as compact continuous states within Large Language Models (LLMs). This approach bypasses the limitations of discrete Chain-of-Thought (CoT) by allowing intermediate reasoning.
Researchers introduce the Focal Modulated Attention Encoder (FATE), a novel transformer architecture designed to accurately forecast multivariate time-series data by explicitly capturing complex spatiotemporal correlations. FATE outperforms existing state-of-the-art methods across
A new vision-language framework, MedReCo-VLM, is introduced to enable entity-aware comparative reasoning in radiology, leveraging large-scale comparative imaging resources to improve clinical follow-up accuracy and analog retrieval.
This research proposes EBiEOT, a new learning paradigm that utilizes inverse entropic optimal transport (OT) and data likelihood maximization to effectively learn conditional distributions in semi-supervised settings, leveraging both paired and unpaired data.
A novel large deviation principle is established for unit norm linear classifiers, demonstrating that all but an exponentially small fraction of such classifiers share nearly identical generalization performance. This highlights an efficiency gap where empirical risk minimization.
Anthropic's Claude Fable 5 model, the first in their Mythos class, is now available within GitHub Copilot, designed for long-horizon, autonomous coding and knowledge-work tasks.
Loryn Brantz, the creator of 'The Good Advice Cupcake,' is protesting Amazon's use of her character for a new AI-animated series without securing her permission.
StepFun has released Step 3.7 Flash, a 198B Mixture-of-Experts (MoE) Vision-Language Model designed specifically for coding agents and search workflows.
NVIDIA has released X-Token, a new projection-guided cross-tokenizer KD method that fixes structural failures in GOLD and significantly improves accuracy on Llama models.
Researchers warn that while AI accelerates code production, the quality of the output may be compromised, posing long-term risks for developers.
A user tested Google's Gemini AI agent by asking it to plan a birthday party, only to discover the AI missed the most important person, highlighting current limitations in real-world personal context understanding.
Tuning and optimizing modern Large Language Model (LLM) serving is challenging due to the myriad of interacting deployment choices, which influence bottlenecks across the entire stack.
The new Opus 4.8 model is now integrated with Dynamic Workflows, a tool for coordinating swarms of subagents.
Large exchanges are creating derivative products around AI tokens, reflecting a growing view of these tokens as raw material inputs rather than mere computational outputs.
OpenAI is expanding trusted access to the GPT-Rosalind model for vetted developers and U.S. government partners to enhance biodefense and public health preparedness through frontier AI.
Introducing Co-ReAct, a novel framework that injects step-level rubrics into ReAct agents to guide their decision-making during inference, leading to more targeted and effective reasoning in complex, multi-step tasks.
This research introduces Oracle-Prompted Policy Optimization (OPPO), a novel reinforcement learning approach that provides per-token credit assignment for LLM reasoning. OPPO leverages Bayesian updates to accumulate oracle signals along a trajectory, yielding token-level advantage
A webinar presentation detailing an end-to-end workflow for designing, training, validating, compressing, and deploying AI-based virtual sensor models onto embedded processors.
This paper introduces Evidence Guided On-Policy Distillation (EDGE-OPD), a novel method for distilling privileged context (like personas or private facts) into Large Language Models. It addresses the challenge that privileged information can degrade general capabilities during tr
We propose the Epistemic Planning Calibration Agentic Workflow (EPC-AW) to mitigate planning failures in LLM-based multi-agent systems caused by misjudging knowledge about plan feasibility. The method improves system-level success by an average of 9.75%.
A new two-stage federated recommendation system that protects sensitive user data by keeping it on the mobile device, only sharing model updates during the training process.
Introducing PathCal, a novel training-free decoding controller that calibrates reasoning paths by distinguishing the distinct functional roles of reflection markers used in Chain-of-Thought generation, leading to improved efficiency and accuracy.
New research introduces Latent Cache Flow (LCF), a method for efficiently communicating between LLM agents by transferring KV caches directly. LCF significantly reduces the size and latency compared to existing methods by jointly compressing keys and values and handling differing contexts.
This research introduces a method to better quantify uncertainty in language model generation by analyzing the geometric trajectories of internal activation updates across layers, showing that these paths reveal where and how errors accumulate.
FusionSense introduces a novel, runtime-adaptive framework that enables energy-constrained autonomous systems to intelligently decide what to compute and transmit at the edge by learning cross-modal dependencies, achieving significant reductions in energy consumption and data transmission
A new method, parallel compaction, is introduced to address context window limitations in long-horizon LLM agents by providing operators with fine-grained control over context summarization and significantly reducing inference time.
This paper introduces S³GNN, a novel method designed to mitigate the Oversqueezing (OSQ) phenomenon in Message-Passing Neural Networks (MPNNs) by efficiently incorporating global information mixing, resulting in significant error reduction and parameter savings.
Gemma 4 models, when paired with multi-token prediction (MTP) drafters utilizing speculative decoding, can achieve up to 3x faster inference speeds without compromising output quality.
Microsoft has introduced MDASH, an AI-driven security platform that uses a multi-model agentic system to automate large-scale code auditing and vulnerability discovery across Microsoft software environments.
A step-by-step guide detailing an advanced federated learning experiment comparing FedAvg and FedProx on a non-IID CIFAR-10 dataset, utilizing the NVIDIA FLARE framework.
A tutorial detailing how to construct an advanced workflow leveraging the SuperClaude Framework to structure interactions with the Anthropic API, incorporating concepts like commands, agents, modes, and session memory.
Nemotron-Labs has developed novel Diffusion Language Models designed to drastically accelerate the generation of high-quality text, pushing the boundaries of generative AI speed.
NVIDIA introduces Gated DeltaNet-2, a novel linear attention layer that decouples the erasing and writing processes within the KV cache, achieving state-of-the-art performance across various language modeling and retrieval benchmarks.