A paper presents an automated tensor scheduling technique to efficiently run LLM inference across both CPU and GPU on consumer hardware, optimizing memory usage and latency.
Reverb is an open-source automatic speech recognition system with integrated speaker diarization, optimized for long-form audio. Its Show HN post has 12 points.
A research project demonstrates that post-training open-weight LLMs using supervised fine-tuning and reinforcement learning from verifiable rewards can pass the Swedish Medical Licensing Exam. Shared on Reddit r/MachineLearning.
SearchOS-V1 is a HuggingFace paper (60 upvotes) proposing a framework for robust multi-agent collaboration in open-domain information retrieval, focusing on reliability and coordination.
A security incident report on HuggingFace revealed that an attacker was not constrained by usage policies, while internal forensic analysis was hampered by the platform's own guardrails. The discussion appeared on Reddit r/LocalLLaMA.