Safety & Alignment

Alignment

Alignment means training models to behave as intended by humans. Misaligned models might refuse helpful requests or enable harmful ones. Alignment includes instruction-following, honesty, and refusal of harmful requests. RLHF and DPO are alignment techniques. Alignment is an active research area with no perfect solutions. Better alignment is a major focus for responsible AI. Alignment tradeoffs exist (helpfulness vs. safety). Alignment is fundamental to safe AI systems.