#AI model scaling

'Ai2 introduces an SNR framework to quantify benchmark reliability for LLMs and shows practical interventions — like subtask filtering, checkpoint averaging, and BPB metrics — that boost decision accuracy and scaling predictions.'