PROM: Pivoted and Regulated Optimization for Multilingual Instruction Learning

Jaeseong Lee, Seung-won Hwang, Hojin Lee, Yunju Bak, Changmin LeeNAACL 2025 · DOI

Introduction

Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their performance remains heavily skewed toward English. For many of the world's languages, adapting these English-centric models typically involves fine-tuning on translated instruction sets. However, this process often encounters a significant hurdle known as negative interference. Instead of benefiting from additional data, the optimization gradients for English and the target language often conflict, leading to suboptimal performance compared to training on less data.

To address this, researchers have introduced PROM (Pivoted and Regulated Optimization), a novel framework that leverages the model's internal "thinking" process to enhance multilingual learning without the typical performance trade-offs.

The English Pivot Hypothesis

A core insight of the PROM paper is that English-centric models like LLaMA2 often use English as an internal "pivot" language. When prompted in a non-English language, the model tends to process semantic information in English within its middle layers before translating that internal representation into the target language at the output layers.

Internal token probability across layers for various target languages, showing English (en) peaking in middle layers.

As shown in the figure above, token probability plots reveal that for a wide variety of languages—including Greek (el), Marathi (mr), and Thai (th)—the probability of English tokens peaks significantly in the middle layers (around layer 20-25) before the model converges on the target language in the final layers. This confirms that the model naturally "thinks" in English before generating a response in the target language.

The PROM Framework

PROM introduces two primary mechanisms to exploit this internal behavior while mitigating gradient conflict:

1. Pivoted Objective

Rather than applying a single loss function at the final layer for both languages, PROM disentangles the training objectives. It applies a semantic loss at a middle layer ($n$) using English data to sharpen the model's core understanding. Simultaneously, it applies a generation loss at the final layer ($N$) using target language data to ensure linguistic fluency and accuracy. This separation allows the model to refine its internal semantic understanding in its pivot language while ensuring the final output is correct in the target language.

2. Regulated Optimization

The strength of the English pivot behavior is not universal; it varies significantly across languages. For instance, the model may pivot strongly for Greek but much less so for Arabic. PROM includes a "regulator" that dynamically assesses the degree of pivoting for a specific language. If the model does not naturally pivot, the system automatically adjusts the training weights to prevent the English semantic objective from over-interfering with the target language generation.

Experimental Evaluation

The researchers evaluated PROM across nine diverse languages (including Arabic, Bengali, Swahili, and Tamil) using major benchmarks such as XQuAD, MLQA, and IndicQA. The results demonstrate that PROM consistently outperforms standard fine-tuning approaches.

Performance comparison on XQuAD, MLQA, and IndicQA benchmarks (EM/F1 scores).

As illustrated in the table, PROM achieved an approximate 50% improvement in Exact Match (EM) scores compared to the base LLaMA2 model. Notably, it outperformed the xLLAMA2(t+e) baseline—which simply appends English data to translated data—proving that PROM's structured optimization effectively resolves the problem of negative interference. Beyond question-answering, PROM also showed significant gains in mathematical reasoning; on the MGSM benchmark, PROM reached an average accuracy of 3.6, compared to 1.0 for the base model.

Conclusion

PROM represents a significant step forward in multilingual LLM adaptation. By recognizing and regulating the internal role of English as a pivot, the framework allows for more efficient and effective fine-tuning across a global array of languages. While the current study focused on LLaMA2-7B, the authors suggest that applying PROM to a wider variety of architectures could reveal if this pivoted behavior is a universal trait of transformer-based models.

Reference: Lee, J., et al. (2025). PROM: Pivoted and Regulated Optimization for Multilingual Instruction Learning. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the ACL.

Summary updated

BibTeX
@inproceedings{lee2025prom,
  author = {Jaeseong Lee and Seung-won Hwang and Hojin Lee and Yunju Bak and Changmin Lee},
  title = {{PROM: Pivoted and Regulated Optimization for Multilingual Instruction Learning}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
  doi = {10.18653/v1/2025.naacl-short.19},
}
Figure