Title: Decoding LLM Mechanics in Multi-Target Cross-Lingual Summarization

Abstract: As audiences increasingly consume information across a variety of linguistic landscapes, multi-target cross-lingual text summarization (MTXLS)—the process of condensing a source document into several target languages—has gained significant relevance. Despite its growing importance, this area remains largely under-researched. To bridge this gap, we present multi-target cross-lingual element-aware (MEA), a novel MTXLS benchmark that encompasses 24 target languages. By evaluating both pipeline and end-to-end methodologies on various large language models (LLMs), we demonstrate that performance in MTXLS continues to trail significantly behind English-only monolingual summarization. To gain deeper insights into how LLMs handle MTXLS, we developed a layer-wise analytical framework to examine the internal mechanics of these models. Our investigation reveals that summarization and translation capabilities do not operate as separate, sequential stages; instead, they emerge concurrently within the model’s deeper layers. These later layers are also where the majority of task-specific processing takes place and where errors frequently originate. Building on these observations, we propose an inference-time activation steering technique that utilizes hidden representations derived from English summarization tasks to direct MTXLS output generation. Empirical results indicate that this approach consistently enhances the quality of summaries across all tested target languages.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC