NAVER LABS re-implements IWSLT 2026 instruction-following pipeline with SeamlessM4T-v2-large and Qwen3-4B-Instruct

NAVER LABS has re-implemented its pipeline for the IWSLT 2026 Shared Task's constrained condition short audio track, utilizing SeamlessM4T-v2-large as the speech encoder and Qwen3-4B-Instruct as the LLM backbone. The team preserved the original three-stage approach of projector alignment, text-only LoRA pre-training, and multimodal merging while constructing 100k synthetic instruction-following examples across ten task types.

• The system adapts to mandated components for the IWSLT 2026 shared task.
• It generates 100k synthetic examples (10k per task) from provided corpora for Stage 3 fine-tuning.
• The primary model achieves a COMET score of 0.781 on EN-ZH speech translation.
• It reaches a BERTScore-F1 of 0.346 on English SQA within the MCIF benchmark.

This re-implementation demonstrates the effectiveness of the adapted pipeline in achieving strong performance metrics on the specified benchmarks.