SyRuP: Enhancing System-Prompt Following via Reward-Guided Prediction in LLM Decoding

SyRuP is an inference-time framework for improving system-prompt adherence while keeping the base language model frozen. It trains a cross-attention reward head on system-prompt-conditioned preference pairs, using the system prompt as a separate memory to score candidate tokens. During decoding, the method reranks the base model’s top-k candidates by combining base logits with the learned adherence reward and an optional contrastive signal based on system-induced logit shifts. The abstract reports consistent gains over prompting and decoding-time baselines with moderate inference overhead, but provides no detailed model, dataset, or numerical results.