MPLUM: A LLM-Assisted Multi-Objective Prompt Learning for High-Quality English-Urdu Summarization
Despite the impressive strides of large language models (LLMs) in text summarization, their effectiveness still hinges on well-crafted prompts, a process that remains labor-intensive, subjective, and difficult to scale. This challenge is magnified for morphologically rich Indo-Aryan languages like Urdu, where English-centric prompting techniques often fall short. To address this, we present MPLUM, a multi-objective evolutionary framework for prompt learning that automates prompt optimization for both English and Urdu summarization tasks. To the best of our knowledge, this is the first attempt of its kind to optimize prompts by simultaneously maximizing summary quality and grammatical correctness, a dual-objective method for producing coherent and readable outputs. Leveraging NSGA-II, our approach evolves prompts using LLM-guided phrase-level mutations: paraphrasing, addition, deletion, and swapping, driven by two fitness/objective functions. The first objective combines lexical and contextual similarity-based relevancy of the sentences in the input document, and the second objective evaluate the grammaticality score using LLM. We introduce language-specific adaptations: English benefits from both crossover and mutation, while Urdu uses mutation-only to better handle morphological complexity. MPLUM is evaluated using tasks from the Natural Instructions collection built on XSUM and XLSUM datasets, across three scenarios: refining strong prompts (Good To Good), improving weak ones (Worst To Good), and optimizing diverse sets (Multi-Good). The optimization process in MPLUM is powered by a mix of open-source and freemium LLMs, including Llama 3.1, Qwen, Mistral, Bloomz, and Gemini 2.0-Flash. The results obtained shows improvements in ROUGE-L F1 score up to 2.68 points for English with Llama 3.1 and 12.09 for Urdu with Gemini 2.0-Flash, with grammar accuracy reaching up to 100%. Llama 3.1 performs best in English; Gemini 2.0-Flash dominates in Urdu. Further, human and LLM-as-a-judge evaluations confirm MPLUM’s robust multilingual performance and scalability. The code of MPLUM is available at our GitHub repository.
Paper
Full text
MPLUM: A LLM-Assisted Multi-Objective Prompt Learning for High-Quality English-Urdu Summarization
Semantic Scholar · Computer Science · 2026
Abstract
Despite the impressive strides of large language models (LLMs) in text summarization, their effectiveness still hinges on well-crafted prompts, a process that remains labor-intensive, subjective, and difficult to scale. This challenge is magnified for morphologically rich Indo-Aryan languages like Urdu, where English-centric prompting techniques often fall short. To address this, we present MPLUM, a multi-objective evolutionary framework for prompt learning that automates prompt optimization for both English and Urdu summarization tasks. To the best of our knowledge, this is the first attempt of its kind to optimize prompts by simultaneously maximizing summary quality and grammatical correctness, a dual-objective method for producing coherent and readable outputs. Leveraging NSGA-II, our approach evolves prompts using LLM-guided phrase-level mutations: paraphrasing, addition, deletion, and swapping, driven by two fitness/objective functions. The first objective combines lexical and contextual similarity-based relevancy of the sentences in the input document, and the second objective evaluate the grammaticality score using LLM. We introduce language-specific adaptations: English benefits from both crossover and mutation, while Urdu uses mutation-only to better handle morphological complexity. MPLUM is evaluated using tasks from the Natural Instructions collection built on XSUM and XLSUM datasets, across three scenarios: refining strong prompts (Good To Good), improving weak ones (Worst To Good), and optimizing diverse sets (Multi-Good). The optimization process in MPLUM is powered by a mix of open-source and freemium LLMs, including Llama 3.1, Qwen, Mistral, Bloomz, and Gemini 2.0-Flash. The results obtained shows improvements in ROUGE-L F1 score up to 2.68 points for English with Llama 3.1 and 12.09 for Urdu with Gemini 2.0-Flash, with grammar accuracy reaching up to 100%. Llama 3.1 performs best in English; Gemini 2.0-Flash dominates in Urdu. Further, human and LLM-as-a-judge evaluations confirm MPLUM’s robust multilingual performance and scalability. The code of MPLUM is available at our GitHub repository.