Researchers have re-implemented the NAVER LABS instruction-following pipeline for the IWSLT 2026 Shared Task, adapting it to use SeamlessM4T-v2-large as the speech encoder and Qwen3-4B-Instruct as the LLM backbone. This adaptation preserves the original three-stage approach of projector alignment, LoRA pre-training, and multimodal merging. To enhance the model, they generated 100,000 synthetic instruction-following examples across ten speech-centric task types. The primary model achieved a COMET score of 0.781 for EN-ZH speech translation and a BERTScore-F1 of 0.346 for English SQA on the MCIF benchmark. AI
IMPACT This research demonstrates adaptation of existing models for specific speech-centric tasks, potentially improving performance on instruction-following benchmarks.
RANK_REASON The cluster describes a research paper detailing the re-implementation of an existing system for a specific task and benchmark.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
