FuseReg proposes regularizing layer fusion to mitigate the reconstruction-generation gap in representation autoencoders, topping HuggingFace daily papers at 112 upvotes. It addresses a known failure mode in latent representation models.
InternW0-Δ is a world action model bridging predictive dynamics and action, trained with over 20,000 hours of open data. It targets embodied and world-model research from Shanghai AI Lab's InternLM line.
TensorFold provides fast, exact LLM decoding on Apple Silicon using MLX, exposed behind an OpenAI-compatible endpoint. It gained 160 stars in a day.
NVIDIA released OpenShell, an open-source sandbox enforcing real runtime limits for local and open agents rather than relying on prompt-based rules. Over 100 firms reportedly joined the safety stack; OpenAI did not participate.
SkillOpt is a text-space optimizer that improves frozen LLM agents via trajectory-driven edits, validation-gated updates, and deployable best_skill.md artifacts. It enables capability improvements without model fine-tuning.