Inference Efficiency of Multimodal vs. Text-Only Models in Zero-Shot Multilingual VQA

Intermediate-task training---fine-tuning a pretrained model on an intermediate task before fine-tuning again on the target task---often improves model performance substantially on language understanding tasks in monolingual English settings. We investigate whether English intermediate-task training is still helpful on non-English target tasks. Using nine intermediate language-understanding tasks, we evaluate intermediate-task transfer in a zero-shot cross-lingual setting on the XTREME benchmark. We see large improvements from intermediate training on the BUCC and Tatoeba sentence retrieval tas

Research goal: How does the inference efficiency (e.g., tokens per second on A100 GPUs) of multimodal models trained with English intermediate-task training (e.g., CLIP + XTREME-R) compare to text-only models on zero-shot multilingual visual question answering benchmarks like MQA?

Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 8.5/10.