DIRECT: Direct Decoding for Efficient and Aligned Sequence Labeling with Large Language Models

2026-07-29

AI Summary

This work addresses the challenges of insufficient domain alignment and low inference efficiency that large language models often encounter in sequence labeling tasks. The authors propose a training-inference co-design framework that integrates Direct Preference Optimization (DPO) with constrained decoding. During training, DPO enhances task alignment by leveraging preference signals, while at inference time, a template-filling mechanism generates only label tokens, reuses prefix key-value (KV) caches to minimize redundant computation, and employs constrained decoding to enforce output format compliance and restrict the candidate label set. Evaluated across eight benchmark datasets, the proposed method significantly outperforms existing approaches, achieving high accuracy while substantially improving inference efficiency.

Abstract

Sequence labeling is a fine-grained information extraction task, yet existing large language model-based approaches suffer from insufficient domain alignment and low inference efficiency. To address these issues, we propose DIRECT, a framework that addresses these issues through training-time optimization and inference-time rectification. Specifically, DIRECT performs Direct Preference Optimization (DPO) after supervised fine-tuning to strengthen task alignment with human preferences, and introduces a controlled decoding process that enforces fixed output formats and restricts predictions to candidate sets. To further improve efficiency, a template-filling mechanism requires the model to generate only label tokens while reusing prefixed content through the KV Cache, thus reducing redundant computation. Experimental results on eight datasets demonstrate that DIRECT achieves significant improvements in both performance and efficiency compared to existing methods.

Problem

• sequence labeling
• domain alignment
• inference efficiency
• large language models

Innovation

• Direct Preference Optimization
• Controlled Decoding
• KV Cache
• Template-Filling
• Sequence Labeling