ByteDance study finds that asking LMMs questions beats making it transcribe text for long document training

The ByteDance Seed model, featuring 7 billion parameters, has demonstrated superior reliability in answering questions on long, image-heavy documents compared to larger models. It was tested on documents four times longer than its training data, showcasing its effectiveness in handling complex information. Instead of transcribing text, the model learns by answering questions and identifying relevant passages independently. This approach marks a shift in how large language models are trained, emphasizing comprehension over rote repetition.