💬 NLP

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

The paper proposes DIRECT, a framework that enhances the performance and efficiency of large language models for sequence labeling by combining Direct Preference Optimization for better alignment with a controlled, template-filling decoding mechanism that minimizes redundant computation.

Original authors: Yilei Wang, Jiaxin Gan, Kexuan Zhang, Ling Li, Wentao Zhang, Peichao Lai

Published 2026-07-30

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Yilei Wang, Jiaxin Gan, Kexuan Zhang, Ling Li, Wentao Zhang, Peichao Lai

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a super-smart robot that has read almost every book on the internet. It can write poems, solve math problems, and chat like a human. This is what scientists call a "Large Language Model" (LLM). But here's the catch: while this robot is a genius at general conversation, it can be a bit of a mess when you ask it to do very specific, rule-heavy jobs. One of these jobs is "sequence labeling." Think of it like a game of "tag" where you have to go through a sentence and slap a specific sticker on every single word. Is "Apple" a company? Is "Tim" a person? Is "March" a date? The robot needs to get every single sticker right, in the exact right order, with no extra words or missing tags.

The problem is that these robots are used to being creative and free-flowing. When you ask them to follow strict rules, they sometimes get confused, forget the format, or just make things up. Also, because they are so chatty, they take a long time to think through every single word, which makes them slow and expensive to run. Scientists have been trying to teach these robots to be better at following rules and faster at doing their work, but it's been like trying to teach a golden retriever to perform precise surgery.

This is where a new method called DIRECT comes in. The researchers behind this paper wanted to fix two big headaches: the robot not listening to instructions well enough, and the robot being too slow. They didn't just try to teach the robot harder; they changed how the robot learns and how it thinks.

First, they gave the robot a special kind of training called Direct Preference Optimization (DPO). Imagine you are teaching a dog to fetch. Instead of just saying "Go get the ball," you show the dog two different ways it could fetch: one where it brings the ball perfectly, and one where it drops it halfway. You then tell the dog, "I like the first one much better." By showing the robot thousands of these "good vs. bad" examples, it learns to understand exactly what humans prefer, making it much more likely to get the job right the first time.

Second, and this is the really clever part, they changed how the robot writes its answer during the actual test. Usually, a robot has to write out the whole sentence again, word by word, including the words it already knows. It's like asking a student to rewrite the whole story of "Cinderella" just to change the color of the dress at the end. DIRECT stops this waste. Instead of rewriting the whole story, the robot is given a "fill-in-the-blank" template. It only has to write the specific stickers (the labels) and skip the rest. The computer remembers the rest of the story instantly, saving a massive amount of time and energy.

The results of this new approach are quite impressive. The researchers tested DIRECT on eight different datasets, which are like different sets of practice problems involving things like finding names of people, places, and organizations. They found that DIRECT was not only more accurate than other top methods but also significantly faster. In fact, on some tests, it was up to 9 times faster than the best existing methods.

When they looked closely at why it worked so well, they found that both parts of their plan were essential. If they removed the special training (DPO), the robot made more mistakes. If they removed the "fill-in-the-blank" speed trick, the robot got the right answers but took forever to do it. By combining strict rule-following training with a smart way to skip unnecessary work, DIRECT managed to get the best of both worlds: a robot that is both a perfectionist and a speedster.

The paper shows that with the right training and a little bit of engineering trickery, we can make these powerful AI tools much more useful for tasks that require precision. It's not just about making the robot smarter; it's about teaching it how to focus and how to work efficiently, turning a chatty genius into a reliable, fast worker.

Technical Summary: DIRECT for Efficient and Aligned Sequence Labeling with LLMs

1. Problem Statement

Sequence labeling, a fine-grained information extraction task encompassing Named Entity Recognition (NER), word segmentation, and Part-of-Speech (POS) tagging, is critical for downstream NLP applications. While Large Language Models (LLMs) have shown promise in low-resource scenarios through fine-tuning, existing approaches face two primary challenges:

Insufficient Domain Alignment: Current methods rely on instruction-based formatting during Supervised Fine-Tuning (SFT). However, this does not guarantee that the model outputs strictly adhere to the expected sequence labeling structures, leading to format violations and misalignments with human preferences.
Low Inference Efficiency: Most approaches retain the standard autoregressive decoding process during inference. This requires the model to generate every token in the output sequence, including structural elements (e.g., parentheses, separators) and known context, resulting in redundant computation and slow inference speeds.

2. Methodology: DIRECT

The authors propose DIRECT (DPO-based Inference RECTification), a framework designed to address alignment and efficiency through a dual mechanism of training-time optimization and inference-time rectification.

2.1. Precision-Oriented Model Optimization (Training Phase)

To enhance domain-specific alignment, DIRECT employs Direct Preference Optimization (DPO) following standard SFT.

Data Construction: The SFT-trained model generates candidate answers for input sequences. Preference pairs $(X, O^+, O^-)$ are constructed where $O^+$ is the ground truth and $O^-$ is a "less-preferred" sample.
Selection Strategy: To maximize entropy gain, less-preferred samples are selected based on a specific criterion: high BLEU scores (indicating fluency) but low F1 scores (indicating poor entity extraction accuracy). This forces the model to learn the distinction between fluent but incorrect outputs and accurate, structured outputs.
Optimization: The DPO objective maximizes the log-probability difference between the preferred and less-preferred outputs under the trainable model relative to a fixed reference model, effectively aligning the model's behavior with human preferences for sequence labeling tasks.

2.2. Inference Rectification (Inference Phase)

During inference, DIRECT introduces a controlled decoding process that enforces strict format adherence and maximizes computational efficiency.

Template-Filling Mechanism: Instead of generating the entire output sequence autoregressively, the model is restricted to generating only the label tokens. The structural components (e.g., words, parentheses, separators) are pre-filled via a template.
KV Cache Reuse: The framework leverages the Key-Value (KV) Cache to avoid redundant attention computations.
1. Prefilling: Before predicting a label for a word $x_i$ , the system processes the preceding known sequence (context $C_{t-2}$ ) in a single pass to compute and store its KV Cache.
2. Incremental Generation: The model only computes new key-values for the specific tokens required for the label (e.g., the left parenthesis and the label characters).
3. Candidate Restriction: The decoding space is constrained to a predefined candidate set of label tokens, ensuring syntactic validity.
Result: This approach consolidates multiple attention operations into a single prefill pass for known prefixes, significantly reducing the number of generation steps.

3. Key Contributions

The paper outlines three primary contributions:

Domain-Adaptive Alignment Optimization: By integrating DPO with a refined preference pair selection strategy and introducing controlled interventions during inference, the method enables LLMs to better understand sequence labeling tasks and align more closely with human preferences.
Significant Improvement in Inference Efficiency: The template-filling mechanism, which requires the model to generate only label tokens while reusing the KV Cache for the rest of the sequence, yields up to 9× faster inference compared to state-of-the-art methods.
State-of-the-Art Performance: Extensive experiments on NER and POS tasks under low-resource settings demonstrate that DIRECT achieves superior performance compared to existing baselines.

4. Experimental Results

The authors evaluated DIRECT on eight datasets (four Chinese and four English NER/POS datasets) under low-resource settings ( $K=250, 500, 1000$ samples).

Performance: Using GLM4-9B-Chat as the backbone, DIRECT consistently achieved the best F1-scores across virtually all datasets. When using LLaMA-3.1-8B-Instruct, it generally attained the second-best results.
- Notable improvements were observed on the MIT-Movie dataset, where F1-scores increased by 13.69% to 15.69% over baselines depending on the sample size.
- Average F1 improvements ranged from 0.37% to 14.72% across various datasets.
Ablation Studies: Removing DPO ("w/o DPO") or inference rectification ("w/ SFT" only) resulted in consistent performance degradation. Specifically, on the Taobao and MIT-Movie datasets, removing the full DIRECT pipeline caused F1 drops of up to 20.21%, confirming the necessity of both training and inference components.
Efficiency: Inference time analysis on the CTB6 dataset showed DIRECT required only 32.86 seconds to process 10 sentences, compared to 309.44 seconds for GNER and 1,468.00 seconds for InstructUIE. This represents an approximate 10x speedup over strong baselines.
Case Studies: Qualitative analysis revealed that baselines often suffered from missing outputs, redundant outputs, or format violations (e.g., failing to label specific words or ignoring structural rules). DIRECT demonstrated superior robustness by strictly adhering to the output format and accurately classifying all entities.

5. Significance

The paper claims that DIRECT offers a practical solution for deploying LLMs in sequence labeling tasks, particularly in low-resource environments. By combining DPO to align model behavior with task-specific requirements and a template-filling mechanism to eliminate redundant generation, the method achieves a dual breakthrough: it significantly improves accuracy and alignment while drastically reducing computational costs. The authors position DIRECT as a highly efficient and robust framework that outperforms existing generative and discriminative approaches in both performance and deployability.