Structure-Conditional Minimum Bayes Risk Decoding

TL;DR

The paper tackles the suboptimality of standard Minimum Bayes Risk (MBR) decoding in open-ended generation where latent structure induces multimodality in the outcome space. It introduces three lightweight structure-aware utility adaptations—Utility Cut-off, Clustering, and Structure Embeddings—to bias MBR toward structure-consistent candidates while preserving the decision-theoretic basis $y^* = \arg\max_h \mathbb{E}_{P(Y|x)}[u(h,Y)]$. A curated dataset spanning three latent structures (dialogue act, emotion, and response structure) demonstrates that standard utilities underperform in cluster-optimality (CO) and cluster-optimal rank correlation (CORC), whereas the proposed methods substantially improve these metrics. The approach yields real-world benefits, achieving up to 13.7 percentage point gains in win rates on AlpacaEval and MT-Bench benchmarks, indicating that structure-aware MBR can enhance generation quality in instruction-following and dialogue tasks with higher outcome-space variability. Overall, structure-aware utilities offer a practical, low-overhead path to align MBR decoding with latent structural variation in natural language generation.

Abstract

Minimum Bayes Risk (MBR) decoding has seen renewed interest as an alternative to traditional generation strategies. While MBR has proven effective in machine translation, where the variability of a language model's outcome space is naturally constrained, it may face challenges in more open-ended tasks such as dialogue or instruction-following. We hypothesise that in such settings, applying MBR with standard similarity-based utility functions may result in selecting responses that are broadly representative of the model's distribution, yet sub-optimal with respect to any particular grouping of generations that share an underlying latent structure. In this work, we introduce three lightweight adaptations to the utility function, designed to make MBR more sensitive to structural variability in the outcome space. To test our hypothesis, we curate a dataset capturing three representative types of latent structure: dialogue act, emotion, and response structure (e.g., a sentence, a paragraph, or a list). We further propose two metrics to evaluate the structural optimality of MBR. Our analysis demonstrates that common similarity-based utility functions fall short by these metrics. In contrast, our proposed adaptations considerably improve structural optimality. Finally, we evaluate our approaches on real-world instruction-following benchmarks, AlpacaEval and MT-Bench, and show that increased structural sensitivity improves generation quality by up to 13.7 percentage points in win rate.

Figures (5)

• Figure 1: The choice of utility function can considerably impact the Minimum Bayes Risk optimum. When the outcome space is structured or multimodal, the MBR optimum may settle between modes, landing in a region of low probability. Here, we present a continuous example featuring a bimodal Gaussian distribution and show the MBR optima (dashed vertical lines) of two utility functions with markedly different behaviours.
• Figure 2: Cluster Optimality and Cluster-Optimal Rank Correlation on the constructed outcome spaces of \ref{['subsec:structures']}. We compare standard BERTScore and BLEURT MBR with the three adaptations to the utility functions proposed in \ref{['sec:structure-conditional-mbr']}.
• Figure 3: System prompt used, for all decoding methods, when generating for AlpacaEval and MT-Bench.
• Figure 4: Examples from the dataset introduced in \ref{['subsec:structures']}. The table contrasts the solution selected by standard MBR (BERTScore-based utility) with the cluster-optimal MBR solution (BERTScore-based) for the given underlying structure. The last example was abbreviated with "…" to fit within the table.
• Figure :