Overfitting Mitigation via Singular Value Decomposition in Minimum Bayes Risk Decoding
SVD-MBR treats the utility matrix as a noisy signal and keeps only its strongest consensus components.
The paper says standard MBR decoding can overfit the metric it optimizes, raising that score while hurting other evaluation metrics. The proposed method uses a low-rank SVD approximation of the pairwise utility matrix to separate consensus from metric noise. Its experiments report broader gains on generalized metrics, with the strongest denoising effect for neural metrics rather than surface-level ones. ArXiv · AI/CL/LG's note
The paper says standard MBR decoding can overfit the metric it optimizes, raising that score while hurting other evaluation metrics. The proposed method uses a low-rank SVD approximation of the pairwise utility matrix to separate consensus from metric noise. Its experiments report broader gains on generalized metrics, with the strongest denoising effect for neural metrics rather than surface-level ones. ArXiv · AI/CL/LG's note
score 4