LLM Alignment

Training the language model itself to act on the user's intent and on agreed human values, so it is helpful, honest and avoids harm, using methods such as reinforcement learning from human feedback (RLHF). It differs from safety guardrails: alignment changes the model's behaviour from the inside during training, while guardrails are an outer layer that checks inputs and outputs at the moment of use.