Minimum Bayes Risk Decoding

• MBR decoding is a decision-theoretic framework that selects the hypothesis minimizing expected risk based on a task-relevant loss function.
• It employs diverse sampling strategies and pseudo-reference construction to accurately estimate expected utility across large, complex output spaces.
• Algorithmic acceleration techniques such as reference aggregation, centroid clustering, and low-rank matrix completion reduce the computational cost while preserving performance.

Minimum Bayes Risk (MBR) decoding is a decision-theoretic output selection method that chooses the hypothesis with optimal expected utility under an implicit or explicit reference distribution. Unlike maximum-a-posteriori (MAP) or greedy decoding, MBR is structured to directly optimize a task-relevant loss or utility function, often yielding outputs with superior correspondence to human or automated evaluation metrics across diverse generation tasks. Multiple algorithmic variants and acceleration techniques have enabled scalable use in neural text generation, translation, instruction following, and other domains.

1. Formal Definition and Expected Utility Principle

Let $x$ be an input instance and $\mathcal Y$ the (usually exponential) output space. With loss function $L(h,y)$ and conditional model $p(y|x)$, the Bayes risk of hypothesis $h$ is defined as

$$R(h) = \mathbb E_{y \sim p(\cdot|x)} [L(h, y)] = \sum_{y \in \mathcal Y} p(y|x) L(h, y)$$

Equivalently, for utility $u(h, y) = -L(h, y)$, the objective is

$$U(h) = \mathbb E_{y \sim p(\cdot|x)}[u(h, y)] = \sum_{y \in \mathcal Y} p(y|x) u(h, y)$$

MBR decoding selects $\hat h$ by

$$\hat h_{\mathrm{MBR}} = \arg\min_{h \in \mathcal Y} R(h) = \arg\max_{h \in \mathcal Y} U(h)$$

In practice, $\mathcal Y$ is intractable, and both candidate hypotheses $H = \{h_1, \ldots, h_N\}$ and pseudo-references $R = \{y_1, \ldots, y_N\}$ are constructed via (pseudo-)sampling or beam search. The empirical estimator is

$$\hat h = \arg\max_{h \in H} \frac{1}{|R|}\sum_{r \in R} u(h, r)$$

or, for model-based weighting,

$$\hat h = \arg\max_{h \in H} \sum_{r \in R} p(r|x) u(h, r)$$

This framework unifies diverse tasks, metrics, and generation strategies, enabling direct optimization of criteria relevant to the end evaluation (Bertsch et al., 2023).

2. Theoretical Properties and Convergence

MBR decoding's empirical effectiveness has motivated deeper analysis of its statistical properties. Under classical Monte Carlo estimation, the MBR selection error converges at rate $O(n^{-1/2})$ in the reference set size $n$, even when $|\mathcal Y| \gg n$ (Ichihara et al., 18 Feb 2025, Bertsch et al., 2023). That is, as the number of samples grows, the approximate solution approaches the Bayes-optimal hypothesis with high probability. This result justifies MBR's robustness in high-dimensional output spaces, clarifying its empirical success.

Comparisons with MAP decoding reveal that MBR typically converges to the optimal solution more rapidly under well-specified loss/utility functions, especially in settings with large output variance (e.g., diverse translation or summarization) (Ichihara et al., 18 Feb 2025). Error decompositions expose the roles of estimator bias—primarily from mismatch between $u(h, y)$ and human utility—and diversity, i.e., the variability in the pseudo-references. Maximizing diversity and minimizing utility bias underlie effective inference scaling laws for MBR (Kamigaito et al., 19 Oct 2024).

3. Sampling Strategies and Risk Estimation

The quality and diversity of the candidate and pseudo-reference sets $H, R$ are critical for the accuracy of MBR estimates. Epsilon-sampling, nucleus sampling, and ancestral sampling offer trade-offs between exploration and sample quality (Freitag et al., 2023, Ohashi et al., 31 Mar 2024). Empirically, epsilon-sampling ($\epsilon \approx 0.02$) yields lower mean human error and larger coverage of plausible translations than naive ancestral or nucleus sampling for neural machine translation, outperforming beam search and other sampling-based decodings in human evaluations (Freitag et al., 2023).

Anomaly detection techniques (e.g., Mahalanobis distance, kNN, LOF) have quantitatively linked MBR performance improvements to the degree of sampling diversity and approximation to the human reference distribution, providing practical proxies for tuning (Ohashi et al., 31 Mar 2024). Increasing the number of pseudo-references reduces estimation variance, but with diminishing returns: performance gains scale approximately as $O(1/N)$ in sample size (Kamigaito et al., 19 Oct 2024).

4. Utility Functions and Metric Bias

MBR decoding is highly sensitive to the utility metric $u(h, y)$. Early applications leveraged simple edit, BLEU, or ROUGE-based metrics, but neural metrics—COMET, BLEURT, BERTScore—now achieve much higher correlation with human judgment. However, optimizing for a single automatic metric induces "metric bias": MBR decoding with metric $m$ substantially increases $m$'s own score, but often produces only marginal or no gain as measured by human ratings. This behavior is reproducible across metrics and language pairs, and transferring evaluation to any correlated neural metric also overestimates improvement (Kovacs et al., 5 Nov 2024).

To mitigate this, ensemble MBR decoders aggregate across several utility functions (e.g., rank average or expected-score averaging over multiple neural metrics), robustly increasing human-aligned quality and eliminating metric reward hacking. Empirically, ensemble-MBR outperforms both greedy and single-metric MBR in total MQM error and fluency/accuracy sub-categories (Kovacs et al., 5 Nov 2024). Practical recommendations are to avoid using the same metric for decoding and evaluation, and to employ metric ensembles for both robust MBR selection and evaluation.

5. Algorithmic Acceleration: Sub-Quadratic MBR

Vanilla MBR decoding incurs $O(N^2)$ utility calls for $N$ candidates with pairwise scoring, which is prohibitive for large $N$ or expensive neural metrics. Several acceleration methods address this bottleneck:

• Reference Aggregation: Aggregate pseudo-reference representations (e.g., $n$-gram vectors, sentence embeddings) into a single centroid or "super-reference", replacing $O(N^2)$ scoring with $O(N)$ while maintaining ≥95% of standard MBR's quality gain in translation tasks (Vamvas et al., 6 Feb 2024).
• Centroid-Based MBR (CBMBR): Cluster reference embeddings into $K \ll N$ centroids (e.g., via k-means on COMET embeddings), and approximate expected utility by evaluating candidates against cluster centroids. This reduces scoring cost to $O(NK)$ and can sometimes even improve translation quality due to better representation of multimodal output spaces (Deguchi et al., 17 Feb 2024).
• Low-Rank Matrix Completion / PMBR: Model the $N \times N$ utility matrix as low-rank, evaluate only a random fraction of entries, and recover the remainder via Alternating Least Squares. This achieves up to $16\times$ reduction in utility evaluations with negligible quality loss (≤0.1 COMET), as the utility matrix is empirically low-rank in text generation tasks (Trabelsi et al., 5 Jun 2024, Natsumi et al., 1 Dec 2025). Agreement-constrained extensions leverage auxiliary metrics for better imputation (Natsumi et al., 1 Dec 2025).
• Confidence-Based Pruning and Sequential Halving: Iteratively prune low-utility candidates using bootstrap confidence intervals (Cheng et al., 2023), or apply medoid identification via Correlated Sequential Halving for a hyperparameter-free approximate MBR with theoretical correctness guarantees and $O(N \log N)$ utility calls (Jinnai et al., 5 Jan 2024).

The mbrs library (Deguchi et al., 8 Aug 2024) provides modular implementations for these algorithmic variants, supporting metrics, expectation estimation, and extensible decoder interfaces.

6. Extensions: Structure Awareness, Out-of-Domain, and Diversity

Recent research broadens MBR decoding to complex contexts:

• Structure-Conditional MBR: Standard similarity-based utility functions can yield poor performance in multi-modal or open-ended tasks (e.g., dialogue, instruction following), where response clusters differ in latent structure (dialogue act, emotion, format). Additions such as act-aware, emotion-aware, and response-type-aware utilities restrict utility computation within structure-consistent groups, producing large improvements (up to 13.7pp win-rate) on instruction-following benchmarks (Eikema et al., 23 Oct 2025).
• Case-Based Decision-Theoretic (CBDT) Decoding: To overcome MBR's reliance on model-sampled pseudo-references (which encode only model knowledge), CBDT uses an out-of-domain memory of reference-evaluated examples. An MBR-CBDT hybrid yields additive gains and greater domain robustness in translation and image captioning (Deguchi et al., 16 Sep 2025).
• Diversity-Promoting MBR (DMBR/KMBR): Instead of selecting a single output, DMBR and $k$-Medoids MBR extend MBR to batch selection, jointly optimizing expected quality and diversity through pairwise penalties or clustering. Compared to diverse beam search and standard sampling, DMBR/KMBR achieve Pareto-dominant quality-diversity trade-offs across MT, summarization, and image/text generation (Jinnai et al., 10 Jan 2024).

7. Empirical Evidence and Applications

MBR decoding yields consistent metric and human-evaluation improvements across NMT, summarization, image captioning, instruction-following, code generation, and grammatical error correction (Bertsch et al., 2023, Wu et al., 3 Oct 2024, Jinnai et al., 10 Jan 2024, Raina et al., 2023). Task-specific metrics (e.g., $F_{0.5}$-score in GEC) can be directly optimized in the risk function, producing explicit control over output precision/recall tradeoffs (Raina et al., 2023). Modern LLM evaluation strategies increasingly incorporate MBR with learned reference-based LLM or ensemble metrics, producing significant win-rate boosts on standard leaderboards (Wu et al., 3 Oct 2024, Eikema et al., 23 Oct 2025).

Hybrid methods, e.g., preference-distillation by Direct Preference Optimization, allow models fine-tuned on MBR-inferred preferences to equal or surpass explicit MBR at inference, but with lower cost (Yang et al., 2023, Wu et al., 3 Oct 2024).

---

Summary Table: Core Variants and Acceleration Techniques

Variant	Time Complexity	Key Idea
Vanilla MBR	$O(N^2)$	Pairwise utility over sampled set
Reference Aggregation	$O(N)$	Aggregate references to single centroid
CBMBR (Centroid)	$O(NK)$ ($K \ll N$)	Cluster refs in feature space
PMBR (Low-rank)	$O(N^2/r)$	Matrix completion (ALS)
Confidence-Pruned MBR	$O(N \log N)$ (avg)	Bootstrap CI + sequential halving
AMBR (CSH)	$O(N \log N)$	Medoid ID via correlated halving
Structure-Aware MBR	$O(N^2)$ (as base)	Utility restricted to structure groups
Ensemble MBR	$O(M N^2)$	Combine $M$ metrics for utility

---

Minimum Bayes Risk decoding provides a powerful, general framework for output selection in sequence generation, driven by explicit optimization of expected utility. Its mathematical tractability, empirical reliability, and extensibility to diverse metrics and domain requirements have established it as a key tool for high-quality, interpretable, and evaluation-aware decoding in modern NLP systems (Bertsch et al., 2023, Deguchi et al., 8 Aug 2024, Kovacs et al., 5 Nov 2024, Natsumi et al., 1 Dec 2025, Kamigaito et al., 19 Oct 2024).