MBR Decoding

MBR decoding selects a translation, transcript, or summary by estimating its expected utility or loss against other plausible outputs, rather than choosing only the model's most likely sequence. Common approaches sample candidate outputs and compare them with metrics such as edit distance, translation quality, or factual consistency; variants use non-autoregressive sampling, task-aware calibration, or low-rank denoising of pairwise scores. The method is studied in machine translation, speech recognition, and summarization, with evaluations including WMT and LibriSpeech benchmarks. A central tension is that MBR depends on both the quality of its candidates and the reliability of its comparison metric, while generating and scoring many candidates can be costly or encourage metric overfitting.