One recent LLM paper a week, in about 10 minutes. Not a bullet-point recap of the abstract: each post walks through the real mechanism with a worked example, an analogy that actually made it click for me, the honest limitations the authors admit to themselves, and where it connects back to RAG.
Language Models Can Control Their Own Attention
A model that tells the inference engine which parts of the context it can skip, cutting attention cost by up to 52% for almost no accuracy loss.
Teaching an LLM to Write CUDA Kernels 2x Faster Than torch.compile
An agent trained with large-scale RL to write CUDA kernels, beating torch.compile’s own output 96.8% of the time.
BDH-CQ: A Reasoning Model That Never Says a Word While It Thinks
A 150M-parameter model that reasons silently in latent space, learns new tasks from examples with no weight updates, and beats every prior cost-accuracy point on ARC-AGI-1.