# Common Interview and Exam Questions — Multimodal AI

Source: https://www.skillbyai.com/en/multimodal-ai/mm-interview-questions

> Answer the multimodal AI questions that come up most in interviews and university exams, with crisp reasoning.

## Questions you should be able to answer aloud

Practise answering each in under a minute. **Why do ViTs fit multimodal models better than CNNs?** They output a token sequence that plugs into transformers and LLMs. **How does CLIP enable zero-shot classification?** Labels become text prompts in a shared embedding space; the image is assigned to the most similar prompt. **Why is batch size important for contrastive learning?** Other pairs in the batch are the negatives, so more negatives make the task harder and the embeddings better. **How does a LLaVA-style VLM see an image?** Vision encoder, projector into the LLM embedding space, image tokens concatenated with text. **What does classifier-free guidance trade off?** Prompt adherence against diversity and image quality at high scales. **Why is a high-resolution image expensive?** Patch tokens grow with area and attention cost grows faster still. **How do you stop a VLM agent being hijacked by text in an image?** Gate privileged actions with validation and human approval.

## Structure for design questions

For "design a system that answers questions about scanned documents", walk through: inputs and volume, representation (page images, OCR, both), retrieval, model choice (hosted or local), output schema and validation, evaluation set, cost per page, privacy and injection risks. Interviewers reward the checklist more than any one model name.

## Explaining it to a non-engineer

If asked to explain multimodal AI to a manager, try: "It is a model that can look, listen and read at the same time, turning each into the same kind of internal notes so it can connect a photo to a question about it." Simple framing shows real understanding.

## Name the trade-off, not just the technique

A strong answer to "concatenation or cross-attention?" is not one word. Say: concatenation is simpler and keeps fine detail but uses context; cross-attention or a resampler keeps context short, which helps for many images or video.

**Quiz:** An interviewer asks why contrastive training like CLIP's benefits from very large batches. Which answer is best?

- [ ] Large batches reduce the image resolution
- [x] Each pair's negatives are the other items in the batch, so more items give more and harder negatives
- [ ] Large batches let the model skip the text encoder
- [ ] Batch size has no effect on contrastive learning

*Answer:* Each pair's negatives are the other items in the batch, so more items give more and harder negatives. In-batch negatives are the core of the loss; bigger batches supply more of them per step.