Created by Jaime S
Latest AI models, benchmarks, algorithms, and applications across robotics, healthcare, coding
Explore the latest content tracked by AI Innovation Radar
TAMP-Nav lets VLMs navigate by selecting 2D pixels (projected to 3D), triggering selective CoT only at key moments, and compressing memory into Space-Time Indicators, then aligning via GRPO rewards for 66.2% SR on R2R-CE using just 90k trajectories.
Existing datasets focus on single editing operations and cannot handle multiple intents executed jointly in one video.
CoinVE-200K introduces 1080p...
QuantumBind Lab provides AI agents a code interface to execute molecular physics operations, bridging reasoning gaps in drug discovery workflows....
GS-Voxel converts pre-optimized 3D Gaussian Splatting scenes into sparse active voxels without any per-scene fitting, then uses a factorized VAE to...
Two studies reveal how AI agent groups defy single-agent assumptions. A PNAS experiment with LLMs shows group size alone can amplify, create, or...
Three complementary advances tackle core deployment bottlenecks for LLM agents.
Two fresh approaches tackle cost and accessibility head-on for frontier models.
Systematic benchmarks of nine models (22.8M–303M params) show compact vision transformers, Swin, and ConvNeXt matching domain-specific foundation models. This challenges the bigger-is-better assumption for vision tasks.
MathForm retrieves Mathlib definitions upfront then iteratively refines outputs using compiler feedback, producing the 367K-example FormalVerse...
AlphaEvolve pushes the matrix multiplication exponent to ω < 2.371177, improving the prior record of 2.371339 held by co-authors Alman and...
Researchers are mapping AI threats across language and physical systems, revealing concentrated attack points and defense gaps.
Time-series foundation models pretrained on financial data produce biological hallucinations on EEG—outputs that are statistically plausible yet...
Generative AI is rapidly moving into fine-grained, controllable visual creation across 3D, images, and animation.
ByteDance open-sourced DeerFlow 2.0 in July 2026 under an MIT license, an AI agent framework built specifically for complex, long-task processing rather than simple chat interactions.
Advanced LLMs consistently underperform their own single-task competence when forced to allocate a shared computation budget across problem suites,...
Prior audit-repair episodes in context make LLM verifiers markedly more lenient, cutting false alarms 9-25% across 15 model-wording combinations by...
A single Vision Transformer model delivers multi-task predictions—cancer subtype, TP53 mutation status, RNA expression, and survival—directly from...
