Technical Screen: Transformers, Attention, Decoding, RLHF, Evaluation, and Optimization

Context: Assume a modern decoder-only LLM unless stated otherwise. Address each prompt concisely but precisely, highlighting trade-offs and practical considerations.

1. Transformer architecture and self-attention

• Explain the Transformer block and how scaled dot-product self-attention works.
• Analyze computational complexity (time/memory) with respect to sequence length n and hidden size d.
• Explain why multi-head attention helps.

2. Attention types and use cases

• Differentiate self-attention vs. cross-attention.
• Define scaled dot-product attention and note alternatives.
• Explain when each attention type is used.

3. Causal (autoregressive) decoding and masks

• Define causal decoding.
• Show how attention masks enforce causality.

4. Decoding and sampling strategies

• Compare greedy, temperature, top-k, nucleus (top-p), and beam search.
• Explain trade-offs in quality, diversity, and latency; give practical tuning guidance.

5. RL-based fine-tuning of LLMs

• Describe the RLHF pipeline: preference data, reward modeling, PPO (or alternatives), and KL control.
• Discuss common stability challenges and mitigations.

6. Evaluation plan for LLMs

• Propose automatic metrics (e.g., perplexity, accuracy), task-based evaluation, human evaluation, and safety/robustness tests.
• Explain how to avoid data leakage and ensure statistical significance.

7. Optimization techniques for training and inference

• Cover optimizers and LR schedules, mixed precision, gradient checkpointing, parameter-efficient finetuning (e.g., LoRA), and distributed strategies (DP/TP/ZeRO/FSDP).
• Include key inference optimizations (KV cache, quantization, speculative decoding).