Professional AI media and industry service platform
JointLight enables controllable relighting of human videos by fusing geometric priors with asynchronous RGB-irradiance joint modeling in a diffusion transformer, achieving state-of-the-art results on both synthetic and real-world benchmarks.
Harvard, NVIDIA, Berkeley, MIT, and Stanford researchers introduce RAX, a framework that scales online exploration for robot agents, achieving 37.4% average improvement over state-of-the-art on five benchmarks without any human demonstrations, using frozen model parameters and self-generated code skills.
Researchers from Renmin University, Rutgers, NUS, and MBZUAI introduce TabClustPFN, a prior-data fitted network that performs zero-shot clustering on tabular data via a single forward pass, automatically estimating the number of clusters and providing a probability distribution over candidate cluster counts, achieving competitive performance on 44 real-world datasets without per-dataset training.
Meta's Muse personal AI agent topped the US App Store and boosted stock 27%, but analysts debate whether its IM-style interface, virtual-machine architecture, and aggressive real-world integration can sustain momentum amid privacy flaws and 'false need' skepticism.
Inception Labs' Mercury 3 diffusion LLM surpasses GPT-6 Luna and Claude Haiku 5.5 on benchmarks like GPQA Diamond and SWE-Bench, achieves 1,068 tokens/second (8x faster), scales 40x parameters while maintaining speed, offers 1M context window at low cost, and excels in AI agent tasks like web navigation and real-time search.
Researchers from HKU and Tsinghua propose VAST, an offline RL framework that adaptively chooses time horizons for value updates by stitching variable-length trajectory segments, improving long-horizon decision-making across 50 OGBench tasks.
Google Labs head Josh Woodward reveals how Google structures its experimental AI team, evaluates product ideas through three lifecycle stages, and avoids the common 'three-year decay' of corporate innovation labs by maintaining organizational independence and clarifying lab objectives.
EchoChat introduces a structured cognitive framework for empathetic speech dialogue, decomposing empathy into perception, Theory of Mind reasoning, and response generation, trained with Acoustic-Anchored Attention and Step-Decomposed Credit Assignment on the EchoDialogue-400K dataset, achieving leading performance on the EchoEval benchmark.
Google open-sourced TorchTPU, a software stack enabling native PyTorch execution on TPUs via device_map="tpu", achieving 97% numerical parity with GPU across 4,917 models and performance within single-digit percentages of JAX, with integrations across vLLM, SGLang, Hugging Face, and TorchTitan.
HyperQ introduces per-token adaptive quantum circuits via a hypernetwork to augment diffusion language models; with 64 qubits, a 1.1B parameter model surpasses 8B models on six zero-shot tasks, and the quantum circuits are validated on IBM's 156-qubit cluster with minimal performance loss.
