Unlocking Latent Discourse Translation in LLMs Through Quality-Aware Decoding

2025-10-08

AI Summary

Large language models (LLMs) still struggle with discourse-level phenomena—such as pronoun resolution and lexical cohesion—in document-level machine translation. This paper presents the first empirical evidence that LLMs implicitly encode discourse-aware translation knowledge in their internal representations. Building on this finding, we propose Quality-Aware Decoding (QAD), a novel decoding strategy that dynamically activates and leverages this implicit knowledge via context-sensitive analysis and alignment with human preferences. Compared to standard decoding methods, QAD significantly improves semantic richness, discourse coherence, and semantic fidelity of translations. It consistently outperforms mainstream baselines across multiple automatic and human evaluations, achieving translation quality closer to professional human output. Our work establishes a new paradigm for uncovering and harnessing latent discourse capabilities in LLMs, advancing the frontier of high-quality document-level translation.

Technology Category

Machine Learning: Large Multimodal Models (LMMs)Natural Language Processing: (Large) Language ModelsSearch and Optimization: Learning to Search

Application Category

Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsEconomics, Online Markets and Human Computation: LLM based quality controls for crowd workSearch and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

Large language models (LLMs) have emerged as strong contenders in machine translation.Yet, they still struggle to adequately handle discourse phenomena, such as pronoun resolution and lexical cohesion at the document level. In this study, we thoroughly investigate the discourse phenomena performance of LLMs in context-aware translation. We demonstrate that discourse knowledge is encoded within LLMs and propose the use of quality-aware decoding (QAD) to effectively extract this knowledge, showcasing its superiority over other decoding approaches through comprehensive analysis. Furthermore, we illustrate that QAD enhances the semantic richness of translations and aligns them more closely with human preferences.

Problem

• LLMs struggle with discourse phenomena in translation
• Propose quality-aware decoding to extract latent discourse knowledge
• Enhance translation quality and alignment with human preferences

Innovation

• Quality-aware decoding extracts latent discourse knowledge
• Method enhances semantic richness of translations
• Aligns machine translation outputs with human preferences