NAVER LABS has re-implemented its instruction-following system for the IWSLT 2026 Shared Task, announced on July 6, 2026. This adaptation utilizes the SeamlessM4T-v2-large speech encoder and the Qwen3-4B-Instruct as the language model backbone, focusing on short audio tracks.
Details of the NAVER LABS System
The re-implementation preserves a three-stage approach that includes projector alignment, text-only LoRA pre-training, and multimodal merging. This method was originally designed for the IWSLT 2025 task but has been updated for the current iteration.
Additionally, NAVER LABS has constructed 100,000 synthetic instruction-following examples across ten different speech-centric task types, with 10,000 examples per task. This dataset aims to facilitate further fine-tuning in Stage 3 of the project.
Performance Metrics Achieved
The primary model developed by NAVER LABS has shown promising results, achieving a COMET score of 0.781 on English to Chinese (EN-ZH) speech translation. Furthermore, it recorded a BERTScore-F1 of 0.346 on English SQA, evaluated against the MCIF benchmark.
Significance of the IWSLT 2026 Task
The IWSLT (International Workshop on Spoken Language Translation) 2026 task emphasizes the importance of accurate instruction-following capabilities in language models. As such, the updates made by NAVER LABS could significantly enhance the performance of automated translation systems in real-world applications.
- COMET Score: 0.781 on EN-ZH
- BERTScore-F1: 0.346 on English SQA
- Synthetic Examples: 100,000 across ten task types
🤖 This article was rewritten by Feed and Figures' editorial AI from a report originally published by arXiv NLP. Facts and quotes are preserved from the original; the rewrite focuses on clarity and structure. For the unedited original, see the source link below.
