Choose a Decoding Strategy for Five Different Products

Your company ships five features, each using an autoregressive model. For each, choose the decoding strategy (greedy, beam search with a beam size and length penalty, or sampling with temperature, top-k and/or top-p), justify it, and name the main failure mode you would guard against.

1. Machine translation of product descriptions.
2. Extraction of invoice fields into JSON.
3. A chat assistant for customers.
4. Marketing-slogan generation, where the user wants 10 different options.
5. Speech-recognition transcription of call recordings (offline).

Then answer two follow-ups. (a) The logits at one step are (2.0, 1.0, 0.5, -1.0): compute the probabilities at temperature 1 and 0.5, and the token set and renormalized probabilities for top-p = 0.9. (b) Why do raw beam-search scores favour short outputs, and what are two fixes?