Tag
Introduces SyRuP, a decoding-time framework that trains a cross-attention reward head to produce token-level adherence scores for system prompts, improving LLM following of complex prompts without model tuning.