
New AI System Reconstructs Moving Objects From Single Video Angles
Researchers introduce Lift4D, a technique that combines 3D modeling with AI priors to capture dynamic deformations and hidden surfaces from monocular footage.
Papers, breakthroughs, benchmarks, and the long-arc trends shaping artificial intelligence. arXiv highlights and lab announcements, distilled.

Researchers introduce Lift4D, a technique that combines 3D modeling with AI priors to capture dynamic deformations and hidden surfaces from monocular footage.

New training method enables continuous dexterous manipulation during locomotion, moving beyond stop-and-grasp limitations.

New approach generates high-quality grasp datasets without human teleoperation, solving a critical bottleneck in dexterous manipulation research.

As LLMs grow more sophisticated, developers and researchers grapple with systems that resist intuitive understanding and traditional evaluation methods.

Researchers propose a mathematical framework that eliminates learned parameters from attention scoring by grounding tokens in abstract group theory.

Researchers propose G2Rec to improve how recommendation systems understand user behavior at industrial scale.

A five-course curriculum tackles the engineering fundamentals of large language models as technical proficiency becomes essential for modern infrastructure development.

New method lets users precisely transform objects in photos using intuitive 3D box interfaces instead of ambiguous text prompts.

New deterministic algorithm achieves optimal accuracy for multicalibration, eliminating need for randomness in trustworthy AI systems.

Researchers introduce proxy-based distillation method that combines nine different data sources to create more capable AI models from first-person video.

New study reveals how to interpret DiffusionGemma's internal reasoning, addressing a critical challenge for understanding next-generation AI systems.

Researchers introduce TimeProVe, a hybrid system that uses lightweight models to propose answers before invoking expensive AI for verification, making long-form video reasoning practical.

New framework creates geometrically coherent 3D objects with dual meanings from different angles in minutes, not hours.

New system compresses months of manual data work into automated workflows that rival human-level performance.

New reinforcement learning approach mimics human behavior more authentically than traditional methods, advancing AI assistant development.

Researchers reveal that frontier multimodal systems struggle to recall past visual information needed for sequential decision-making tasks.

Researchers introduce an intelligent video understanding system that reasons through footage selectively, matching larger models while using a fraction of the compute.

A bottleneck-shaped model structure cuts computing costs by 22% while improving performance, challenging conventional neural network scaling wisdom.