SyRuP: Enhancing System-Prompt Following via Reward-Guided Prediction in LLM Decoding
2026-07-27 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors present SyRuP, a method to help large language models better follow instructions given in system prompts without changing the model itself. Instead of retraining, SyRuP learns to score how well each word in a response matches the prompt and uses these scores to pick the best response from a set of options. Tests show that SyRuP improves how closely the model sticks to the prompt compared to other simple methods, with only a small extra cost when the model generates text. This approach offers a practical way to guide language models more precisely during text generation.
Large Language ModelsSystem PromptIn-Context LearningDecoding-Time RerankingCross-AttentionToken-Level ScoringPreference PairsInference-Time ControlBase Model FreezingLogit Re-ranking
Authors
Seoyeon Kim, Minjae Kang, Jaehyung Kim
Abstract
Large Language Models (LLMs) are increasingly controlled through system prompts that specify roles, styles, formats, and safety requirements. However, models follow these prompts only implicitly through in-context learning, which can be insufficient for complex or compositional prompts. Existing approaches often require model tuning or response-level reranking, limiting their practicality for lightweight inference-time control. We introduce SyRuP, a decoding-time framework for improving system-prompt adherence while keeping the base LM frozen. SyRuP trains a cross-attention reward head from system-prompt-conditioned preference pairs, treating the system prompt as a separate memory to produce token-level adherence scores. At inference, SyRuP reranks the base LM's top-k candidates by combining base logits with the learned reward signal and an optional contrastive signal capturing system-induced logit shifts. Experiments on system-prompt following benchmarks show that SyRuP consistently outperforms prompting and decoding-time baselines with moderate inference overhead. These results suggest that explicit token-level guidance is an effective and practical mechanism for reliable system-prompt following.