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

2026-07-27 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

SyRuP is a novel decoding-time framework designed to enhance Large Language Model (LLM) adherence to system prompts, particularly for complex or compositional instructions. Unlike methods requiring model tuning or response-level reranking, SyRuP keeps the base LM frozen. It operates by training a cross-attention reward head from system-prompt-conditioned preference pairs, treating the system prompt as a separate memory to generate token-level adherence scores. During inference, SyRuP reranks the base LM's top-k candidate tokens by combining the base logits with its learned reward signal and an optional contrastive signal. Experiments demonstrate that SyRuP consistently outperforms existing prompting and decoding-time baselines, introducing only moderate inference overhead.

Key takeaway

For Machine Learning Engineers deploying Large Language Models with complex or compositional system prompts, SyRuP offers a practical decoding-time solution to improve adherence without costly model fine-tuning. You should consider integrating this framework to achieve more reliable and consistent model outputs, especially for safety or style requirements, given its demonstrated performance improvements over baselines and moderate inference overhead.

Key insights

Explicit token-level guidance via reward-guided prediction significantly improves LLM system-prompt following.

Principles

• In-context learning alone is often insufficient for complex system prompts.
• Explicit token-level guidance offers a practical mechanism for reliable prompt following.

Method

SyRuP trains a cross-attention reward head using system-prompt-conditioned preference pairs to produce token-level adherence scores, then reranks base LM's top-k candidates by combining base logits with this reward and an optional contrastive signal.

In practice

• Enhance LLM system-prompt adherence without requiring base model fine-tuning.
• Achieve lightweight inference-time control for complex or compositional prompts.

Topics

• Large Language Models
• System Prompts
• Decoding Strategies
• Reward Models
• Prompt Following
• Inference Optimization

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Soft-Prompt Tuning for Fair and Efficient LLM Benchmark Evaluation — Soft-prompt tuning efficiently reveals true LLM knowledge by overcoming format-following limitations in benchmarks.
• FLy: A New Paradigm for Speculative Decoding — Accepting Semantically Correct Drafts Beyond Exact Match — FLy accelerates LLM inference by accepting semantically valid drafts, overcoming exact-match limitations without training.
• SODA: Semi On-Policy Black-Box Distillation for Large Language Models — A static snapshot of a student's own inferior responses provides an efficient, targeted contrastive signal for black-box LLM distillation.
• AlignCoder: Aligning Retrieval with Target Intent for Repository-Level Code Completion — AlignCoder enhances RAG for repository-level code completion by using multiple candidate completions to refine queries and training a retriever with reinforcement learning.
• Lossless Prompt Compression via Dictionary-Encoding and In-Context Learning: Enabling Cost-Effective LLM Analysis of Repetitive Data — LLMs can learn dictionary encodings in-context, enabling lossless prompt compression without fine-tuning and preserving analytical accuracy.
• C-GAP: Class-Aware and Online Prompting Improves Vision-Language Models on Imbalanced Classes — Iteratively refining language prompts with an LLM, using composite captions, significantly boosts minority-class detection in frozen vision-language models.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.