Decoding strategies + speculative decoding + JSON mode + DSA Greedy

Interview questions to prep

1. Prove the greedy choice — why is the locally-optimal pick safe globally? (Exchange argument or staying-ahead.)
2. When does greedy fail on a similar-looking problem, and what would you reach for instead (DP, BFS)?
3. Walk through edge cases that often break naive greedy: ties, negatives, single element.

Interview questions to prep

1. Prove the greedy choice — why is the locally-optimal pick safe globally? (Exchange argument or staying-ahead.)
2. When does greedy fail on a similar-looking problem, and what would you reach for instead (DP, BFS)?
3. Walk through edge cases that often break naive greedy: ties, negatives, single element.

Interview questions to prep

1. Compare greedy, beam, top-k, and nucleus (top-p) decoding.
2. Why is beam search usually a bad choice for open-ended generation?
3. What does temperature actually do to the softmax distribution?

Interview questions to prep

1. How does speculative decoding speed up inference without hurting quality?
2. Why doesn't speculative decoding always help — what's the relationship between draft acceptance rate and speedup?

Interview questions to prep

1. How would you force an LLM to emit valid JSON — and what are the failure modes?
2. Compare grammar-constrained decoding vs prompt-+-retry-+-validate — when does each fit?

Interview questions to prep

1. Implement top-k and nucleus sampling inside an autoregressive generate() loop.
2. How do temperature, top-k, and top-p interact when a model becomes repetitive or incoherent?
3. What stopping conditions do you need for a chat model beyond max_new_tokens?