Plan2Align: Predictive Planning Based Test-Time Preference Alignment in Paragraph-Level Machine Translation

2025-02-28

AI Summary

Current paragraph-level machine translation suffers from semantic incoherence and information loss, primarily due to the myopic, token-by-token generation of non-expansive language models and the inability of conventional preference alignment methods to model cross-sentence consistency. To address this, we propose a test-time predictive planning framework—the first to integrate Model Predictive Control (MPC) into paragraph-level preference alignment—enabling global optimization of translations via iterative re-scoring and context-aware re-generation. We instantiate this framework end-to-end atop LLaMA-3.1 8B. Evaluated on the WMT24 literary translation benchmark, our approach significantly improves paragraph-level coherence and faithfulness, matching or surpassing state-of-the-art training- and test-time alignment methods. It establishes a scalable, consistency-aware paradigm for long-text translation.

Technology Category

• Planning, Routing, and Scheduling: Planning with Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Machine Translation (MT) has been predominantly designed for sentence-level translation using transformer-based architectures. While next-token prediction based Large Language Models (LLMs) demonstrate strong capabilities in long-text translation, non-extensive language models often suffer from omissions and semantic inconsistencies when processing paragraphs. Existing preference alignment methods improve sentence-level translation but fail to ensure coherence over extended contexts due to the myopic nature of next-token generation. We introduce Plan2Align, a test-time alignment framework that treats translation as a predictive planning problem, adapting Model Predictive Control to iteratively refine translation outputs. Experiments on WMT24 Discourse-Level Literary Translation show that Plan2Align significantly improves paragraph-level translation, achieving performance surpassing or on par with the existing training-time and test-time alignment methods on LLaMA-3.1 8B.

Problem

• Addresses paragraph-level translation coherence issues
• Improves semantic consistency in long-text translation
• Enhances translation quality using predictive planning

Innovation

• Predictive planning for paragraph-level translation
• Model Predictive Control for iterative refinement
• Test-time alignment framework for coherence