Instruction Following / Alignment

Teaching language models to follow user instructions accurately and safely. CoT later became important for instruction-following systems like InstructGPT and ChatGPT, where showing reasoning steps helps the model follow complex multi-step instructions. RLHF reward models increasingly gave higher scores to outputs with explicit reasoning chains.