Frédéric Berdoz

New inference technique boosts LLM alignment without extra training

Researchers have developed a new inference-time technique called alignment-aware decoding (AAD) to improve the alignment of large language models. AAD operates without requiring additional training beyond standard prefe…