EngramEdit proposes a conditional-memory approach that decouples knowledge updates from base model weights, enabling targeted edits without full fine-tuning.
Tetris3D introduces a method for generating 3D scenes where objects physically fit together, receiving 35 upvotes on HuggingFace Papers.
Alibaba's Qwen organization updated Qwen3-TTS (13.6k stars) and Qwen-Agent (17.1k stars) repositories, signaling continued investment in the Qwen voice and agent stack.
GRACE proposes a generation-aware latent compression method to improve efficiency in video generation diffusion models. It reached 61 upvotes on HuggingFace Papers.
Long-WAM is a new research effort on scaling context windows in world-action models, appearing simultaneously on arXiv and HuggingFace Papers with 86 upvotes.