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.
Introduction: The "Middle Ground" Problem
In natural language generation, Minimum Bayes Risk (MBR) decoding is often employed to find a "consensus" response from a pool of candidates by maximizing average similarity. However, in open-ended tasks like dialogue, the model's distribution is often multimodal—for instance, it might be unsure whether to respond with a concise list or a detailed paragraph.
Standard similarity-based utility functions (e.g., BERTScore) are often "structure-blind." When faced with multiple structural modes, MBR tends to select a "compromise" candidate that sits between these modes. Such a response might awkwardly mix styles, satisfying the mathematical average but failing to provide a high-quality, structurally coherent output.
Proposed Methods
To address this, we propose three lightweight adaptations to the MBR utility function $u(h, r)$:
- Utility Cut-off: A threshold $\tau$ is applied; similarities below $\tau$ are treated as zero. This forces the decoder to ignore structurally distant candidates and focus on local consensus within a specific mode.
- Clustering: The candidate space is partitioned into clusters using sequence embeddings. MBR is then restricted to the largest (dominant) cluster, ensuring the selection of a "pure" representative of a single structural intent.
- Structure Embeddings: The utility function is weighted by the cosine similarity of candidates in a fine-tuned, structure-aware embedding space. This softly biases the selection toward candidates sharing the same latent structure.
Experimental Results
Our analysis confirms that standard MBR is cluster-optimal in fewer than half of the cases in our curated dataset. In contrast, our proposed methods significantly improved "Cluster Optimality" reaching up to 98% for response structures.
On real-world benchmarks, these structural improvements translated directly into higher generation quality. As shown in the table below, the Utility Cut-off method achieved a 90% win rate on MT-Bench (single-turn), a 13.7 percentage point improvement over standard MBR.
Conclusion
By making MBR "structure-aware," we can avoid the "middle ground" trap and select responses that are both representative and structurally coherent. While these methods depend on the quality of underlying embeddings and model calibration, they offer a significant boost in performance for open-ended generation tasks with minimal additional computational overhead.
Summary updated