AI
- Hugging Face Papers — Top paper “LongLive-Plug” (NVIDIA) proposes distilling video-generation capabilities once per model family into reusable LoRA modules that plug into compatible downstream models training-free, instead of re-distilling each one from scratch.
- arXiv — EleutherAI researchers show training-data attribution can identify which fine-tuning examples drive emergent misalignment, letting it be amplified or suppressed — though attribution transfers within a model family but not across families.
- arXiv — Study of verifiable grade-school math finds models often reach correct final answers via reasoning traces that fail step-by-step verification, meaning a correct answer isn’t evidence the reasoning itself was valid.