Original editorial coverage of computer vision, AI perception, and machine recognition — the research and ideas behind what "I See You" actually means for a machine.
Mixture of Probes (MoP) shows that auxiliary visual or audio data, even when unavailable at inference, can boost multimodal LLM performance by up to 65%, reshaping training strategies for AI perception.
StereoSplat+ shows how diffusion can turn a lone stereo rig into a high‑fidelity 3D Gaussian splatting pipeline, boosting novel‑view rendering on KITTI‑360.
A deep dive into PEVA, the new diffusion‑transformer that predicts egocentric video from full‑body motion, and why it matters for embodied AI planning.
LingBot‑VA 2.0 introduces a causal DiT with a sparse‑MoE video stream, 225 Hz asynchronous control and multi‑chunk prediction, pushing robot manipulation toward true physical AI.
Explore NVIDIA’s tile‑based GPU programming with cuTile, Triton, and flash attention, and see why this paradigm is a game‑changer for AI workloads.
Thinking Machines Lab argues that decentralized, fine‑tuned AI—using multimodal interaction and LoRA‑based weight adapters—can keep ownership and alignment in the hands of users.
Anthropic built its new Claude desktop agent, Cowork, in just ten days, letting AI read and edit local files while raising fresh safety questions for responsible AI adoption.
Listen Labs raised $69 M to scale AI‑driven video interviews, tackling fraud and slashing research cycles from weeks to hours while reshaping product development.
A new framework uses mutual information to evaluate and optimise cameras, telescopes and microscopes, cutting compute and memory while matching end‑to‑end performance.
A deep dive into GRASP, the new gradient‑based planner that lifts trajectories, adds stochastic state noise, and reshapes gradients to make long‑horizon world‑model planning practical.
Discover OpenAI's GPT-5.6, a groundbreaking three-tier model family that sets new standards in AI performance, with Sol leading the Artificial Analysis Coding Agent Index at 80
Meta Superintelligence Labs releases Muse Spark 1.1, a multimodal reasoning model with a 1,000,000-token context window for agentic tasks
Discover LingBot-World-Infinity, a 14B causal video generation model that simulates interactive worlds with unprecedented quality and consistency
Countries are investing in AI capabilities to drive innovation and economic growth while ensuring responsible use of technology
Discover how open models are driving AI research forward with NVIDIA's 74 accepted papers at ICML 2026
NVIDIA and Hugging Face collaborate to bring cutting-edge models and frameworks to LeRobot, empowering robotics developers with open-source tools and resources
Google, NYC educators, and industry leaders gathered to discuss AI's role in preparing students for tomorrow's careers
Discover the latest AI updates from Google, including Gemini 3.5 Live Translate and Android 17, designed to make your devices and apps more helpful
Google's Gemini API updates Managed Agents with background execution, remote MCP server integration, and custom function calling
Discover how Goose, a free and open-source AI coding agent, is challenging the expensive AI coding revolution with its local machine capabilities
Railway raises $100 million to take on AWS with AI-native cloud infrastructure, promising faster deployment and lower costs
Google's new search box redesign invites users to think out loud and interact with AI in a more conversational way
Discover how adaptive parallel reasoning is transforming the field of artificial intelligence by enabling models to dynamically allocate compute between parallel and serial operations
The Berkeley Artificial Intelligence Research Lab congratulates its class of 2026 graduates, who have made significant contributions to AI and machine learning.
As AI costs drop, agents will become the dominant workload for data systems, requiring new designs for data systems for, of, and by agents
Google Research introduces SensorFM, a wearable health foundation model pretrained on 1 trillion minutes of sensor data from 5 million people
Learn how to create a T4-friendly autonomous data science agent using DeepAnalyze-8B, sandboxed code execution, and iterative analysis
Discover MuScriptor, a groundbreaking open-weight decoder-only Transformer for multi-instrument music transcription to MIDI, trained on 170k real recordings and 1.45M synthetic MIDIs
NVIDIA Vera leads the charge in max single-threaded CPUs, boosting AI factory revenue and agent performance with unparalleled speed and efficiency
NVIDIA Nemotron 3 Ultra offers leading performance at lower cost with LangChain's Deep Agents harness, achieving highest accuracy among open models
NVIDIA's GeForce NOW cloud gaming service expands to Toronto with new RTX 5080-powered server, bringing high-performance gaming closer to Canadian members
LightCrafter achieves controllable and consistent relighting in videos using a hybrid pipeline that combines physically-based rendering with video diffusion refinement
FedTR, a novel federated learning framework, achieves high accuracy in industrial visual inspection tasks with limited data availability
Discover how LOGOS, a novel transformer-based approach, leverages textual prompts to improve object detection accuracy in complex aerial environments
Discover how ThermoField, a novel framework, revolutionizes thermal imaging by jointly reconstructing geometry and estimating thermophysical properties
Researchers discover a method to deceive attention-based defenses in Vision Transformers, revealing a fundamental limitation in using attention magnitude to detect adversarial attacks
Researchers use low-cost UAV photogrammetry to monitor shoot elongation in deciduous trees, shedding light on climate change's impact
Discover how the GIRAF model is transforming the field of embodied AI and graphics with its innovative approach to synthesizing realistic human interactions with articulated objects
DreamCharacter-1 calibrates pretrained 3D foundation models for high-fidelity character generation, surpassing state-of-the-art methods
Discover how SpaR3D-MoE enables adaptive 3D spatial reasoning from sparse RGB inputs, outperforming existing models on VSI-Bench, ScanQA, and SQA3D benchmarks
Discover how MiLSD, a novel line segment detector, achieves high accuracy under a sub-megabyte budget, paving the way for embedded vision systems
Researchers investigate the connection between counterfactual fairness and group fairness in image classification, revealing a surprising disconnect
Discover the Difficulty-Aware Medical Instructional Video Question Answering challenge, a benchmark for medical video understanding systems
Discover CoFINN, a novel physics-informed deep learning framework that improves aerodynamic force prediction accuracy by up to 34%
Discover how fine-tuned latent diffusion models can generate culturally consistent yet novel Ulos motifs, preserving tradition while meeting contemporary design demands
LipSSD introduces a Lipschitz-constrained approach to object detection, enhancing robustness against adversarial attacks
Researchers propose a pipeline for generating photorealistic defocus datasets with diverse lens characteristics
Discover how generative randomization and cross-variant self-supervised learning can help deep neural networks overcome spurious correlations and achieve robust visual representations
A deep learning pipeline for disease severity quantification in field crops achieves 98.20% pixel accuracy, enabling real-time automated crop monitoring
Discover how SegAnswer, a novel approach to visual reasoning, achieves consistent improvements in multimodal large language models
TRIG framework achieves state-of-the-art performance in pose estimation and 3D reconstruction for autonomous driving
Researchers combine image classification and vessel segmentation for AI-based screening of retinopathy of prematurity, improving detection accuracy in Kenyan preterm infants
Discover how Legato 2, a novel pipeline, is transforming optical music recognition and sheet music understanding with its sequential system-by-system approach
New research highlights the impact of environmental illusions on autonomous driving systems, posing serious safety risks
DeSeG framework achieves state-of-the-art performance in synthesizing physically plausible human-scene interactions
Discover how ARMS, a novel Anchor-Relational Motion Streaming framework, generates temporally continuous and socially coherent human motion from text
New research optimizes the adaptive loop filter in Versatile Video Coding, reducing buffer access and encoding time
Image2Sim constructs high-quality interactive environments from RGB-D image sequences, achieving strong improvements on major benchmarks
Discover how Scene Graph Thinking enhances multimodal large language models with structured visual reasoning
A new diagnostic tool reveals whether learned shortcuts can be restored after unlearning, shedding light on the effectiveness of shortcut-mitigation methods
Introducing SAMPLe, a novel optimizer that improves prompt learning in Vision-Language Models by balancing exploration and exploitation
FASR++ enhances face recognition in low-quality images with a diffusion-model-based super-resolution algorithm
Discover how visual encoding hijacking induces bias in vision models and the impact of chart design on machine learning
Discover REVIVE, a groundbreaking framework that detects and recovers from vandalism-induced occlusion attacks in autonomous vehicles
Discover how Cluster-Guided Vector Quantization improves image compression efficiency by 20% without sacrificing visual quality
Discover how AI-driven vision-language models are transforming wound monitoring and severe adverse event detection in clinical settings
Discover how leveraging disease taxonomy improves accuracy in chest X-ray classification, enabling better diagnosis and treatment
Discover how a novel two-stage diffusion-based framework enhances cloud microstructure resolution, paving the way for improved climate and sustainability research
Discover how generative image models can be harnessed for training-free primitive shape abstraction, enhancing robotics and scene understanding
Discover how 3D facial meshes and hierarchical classification can improve facial phenotyping for clinical diagnosis
Discover how Patch Knowledge Transfer optimizes AI-generated image quality assessment, achieving a 67.7% reduction in computational costs
Discover how a new pretraining framework enables efficient deployment of pathology foundation models on edge devices
Discover how Bayesian 3D Gaussian Splatting is transforming 3D scene reconstruction with native uncertainty and adaptive complexity control
Ground3D-LMM enables fine-grained 3D point grounding and spatial reasoning with real-world measurements
Discover how Light-Omni is transforming video understanding with its innovative reflexive agent framework
Discover how ordinary vision datasets can contain exploitable adversarial surfaces, threatening model security and reliability
Discover a novel approach to gaze estimation using just one camera and one light source, paving the way for more accessible eye tracking
CanvasAgent enables complex image creation and editing
UAV detection and tracking in foggy conditions