Sequence-Centric Stochastic Beam Search

• Sequence-centric SBS is a probabilistic decoding algorithm that samples full candidate sequences without replacement, combining beam search efficiency with controlled randomness.
• It employs methods such as Gumbel-Top-k perturbations and conditional Poisson sampling to ensure diversity and produce low-bias, variance-minimized estimators.
• Empirical evaluations demonstrate that SBS enhances outcomes in tasks like NMT, LLM uncertainty quantification, and protein engineering by enabling robust multi-objective optimization.

Sequence-centric stochastic beam search (SBS) refers to a family of probabilistic decoding algorithms for sequence models that generalize deterministic beam search by introducing randomness via structured sampling—typically without replacement—over full candidate sequences. The approach seeks to ameliorate the deficiencies of both greedy beam search (low diversity, biased statistics) and naïve multinomial sampling (high variance, duplication), generating representative, diverse, and statistically principled sets of sequences. Prominent instantiations leverage conditional Poisson sampling (Meister et al., 2021), Gumbel-Top-$k$ perturbations (Kool et al., 2019), or stochastic ranking procedures. SBS has found diverse applications in natural language processing, uncertainty quantification in LLMs, protein engineering with masked LLMs, and reasoning chain calibration in large models.

1. Foundational Principles and Formulations

Sequence-centric SBS reframes sequence generation as the problem of sampling a set of full hypotheses, of size $K$, from the model distribution $p_\theta(\cdot\mid x)$, without replacement, and typically with probability proportional to each candidate's likelihood. Standard beam search deterministically selects the $K$ highest-updated partial sequences at each token step. In contrast, stochastic beam search replaces this greedy selection with probabilistic subset sampling defined over the expanding beam at each step.

Key instantiations:

• Gumbel-Top-$k$ Trick: Perturb log-probabilities $\phi_i$ of all (potentially exponentially many) sequences by i.i.d. Gumbel noise, then take the top-$k$ perturbed scores. This yields exact samples without replacement from the categorical over sequences (Kool et al., 2019).
• Conditional Poisson Sampling: Select $K$ candidates at each expansion step according to the conditional Poisson (CP) law,

$$Q_t(Y_t\mid Y_{t-1}) = \begin{cases} \dfrac{\prod_{n\in Y_t} w_n}{\sum_{U\subseteq B_t,\,|U|=K} \prod_{n\in U} w_n} & |Y_t|=K \ 0 & \text{otherwise} \end{cases}$$

with weights $w_n$ determined by local (optionally temperature-scaled) conditional likelihoods (Meister et al., 2021).

Both approaches enforce structured sampling without replacement (SWOR) among candidates,, promoting coverage and variance control, and yield unbiased or low-bias estimators for sequence-level metrics.

2. Algorithmic Structure and Computational Aspects

A representative SBS algorithm proceeds in the following stages (see (Kool et al., 2019, Meister et al., 2021)):

1. Initialization: Start with a root beam (e.g., containing only $\langle$BOS$\rangle$).
2. Expansion: At each decoding step $t$, for each partial sequence $y\in Y_{t-1}$, generate all possible one-token extensions, forming $B_t$.
3. Score Assignment and Perturbation:
   • Assign a score $p(y\mid x, y_{<t})$ or logit $\phi(y)$ to each candidate.
   • Optionally, apply temperature scaling or incorporate auxiliary rewards/self-evaluation terms.
   • For Gumbel-based SBS: Add i.i.d. Gumbel noise to each candidate's score.
4. Stochastic Selection:
   • For Gumbel SBS, select the top-$K$ perturbed candidates globally.
   • For Conditional Poisson SBS, draw a random $K$-subset via CP-SWOR, using combinatorially normalized weights (see $O(NK)$ DP for normalization (Meister et al., 2021)).
5. Recursion: Repeat for $T$ token-level expansions, maintaining the beam at size $K$ throughout.
6. Output: Return the final set of $K$ sequences.

Computational complexity per step is $O(K|V|)$ for Gumbel SBS, and $O(K^2|V|)$ for CP-SBS (after DP-based normalizer). Backtracking beam histories, estimation of sequence inclusion probabilities, or multi-objective scoring can be efficiently integrated.

3. Varianced-Minimizing Estimators and Theoretical Guarantees

SBS-generated beams support estimators with favorable variance properties for expected sequence-level functionals:

• Horvitz–Thompson (HT) Estimation: For any $f(y)$,

$$\hat G_{HT} = \sum_{y \in Y_T} \frac{p(y|x)}{\pi(y)} f(y)$$

where $\pi(y)$ is the marginal inclusion probability. This estimator is unbiased (Meister et al., 2021).

• Priority Sampling Estimators (Gumbel-Top-$k$): Incorporate correction factors for SWOR probabilities, yielding unbiased or variance-lowering statistics for quantities like expected BLEU or entropy (Kool et al., 2019).
• Theoretical Bounds: In uncertainty quantification, when the cumulative probability mass of the beam exceeds $1-1/(2\sqrt{M})$, the bias-variance trade-off of the beam-weighted estimator strictly dominates multinomial sampling's MSE (Fadeeva et al., 10 Dec 2025).

SBS avoids the high duplication rate of multinomial sampling on peaked distributions, resulting in lower-variance estimates for almost any reasonable diversity budget.

4. Methodological Extensions and Self-Evaluation Guidance

Advanced SBS variants incorporate model-internal or auxiliary scoring mechanisms:

• Self-Evaluation Guided Decoding: Combines autoregressive model confidence and an LLM-based correctness/confidence score at each step, enabling composite stepwise objectives $$E(s^{1:T}) = \prod_{t=1}^T \mathbb{P}(s^t \mid x, s^{1:t-1})^\lambda \cdot C(s^t)^{1 - \lambda}$$ (Xie et al., 2023). At each expansion, sampling is performed with probabilities proportional to $\exp(E/\tau)$, with an annealed temperature promoting initial exploration and later exploitation.
• Scalarized or Pareto Ranking: For protein design, SBS can accommodate black-box objectives via Pareto non-dominated sorting or smooth Tchebycheff scalarization layered atop the sequence likelihood, enabling flexible multi-objective search (McCarter et al., 11 Mar 2026).

These extensions allow SBS to traverse solution spaces more robustly, guide search away from locally fluent but globally invalid solutions, and support highly calibrated generation under multiple constraints.

5. Practical Implementation and Domain-Specific Adaptations

Across NLP and biomolecular engineering, SBS has been effectively integrated with both autoregressive and masked LLMs:

• In NMT (WMT’14 En–Fr, Transformer), CP-SBS with the HT estimator exhibits lower RMSE in BLEU and entropy estimation at practical $K$ (e.g., $K\lesssim 50$) and returns more diverse outputs than MC sampling or deterministic beams (Meister et al., 2021).
• In LLM uncertainty quantification (QA tasks), sequence-centric SBS achieves higher and more stable prediction-rejection ratios than multinomial or sampling-based consistency measures, especially for short, peaked answer distributions (Fadeeva et al., 10 Dec 2025).
• In masked protein LLMs, SBS with efficient pseudo-perplexity updates and 1-edit neighborhood expansion yields higher biologically relevant hit rates and supports multi-objective optimization (e.g., synthesizability, thermostability), both in silico and in vitro (McCarter et al., 11 Mar 2026).

Efficient dynamic programming, nucleus truncation, Gumbel perturbation batching, and GPU parallelism are employed to maintain tractable runtime and memory. Model-specific peculiarities, such as non-autoregressive scoring in MLMs, are explicitly addressed.

6. Empirical Findings and Performance Benchmarks

Empirical studies across domains consistently indicate that:

• SBS and its advanced forms (CP-SBS, beam-weighted estimators) exhibit lower variance and/or lower bias for sequence-level metrics compared to both standard beam search and multinomial sampling—with especially large gains as the generative distribution sharpens (i.e., lower temperature, peaked solutions) (Meister et al., 2021, Fadeeva et al., 10 Dec 2025).
• In multi-step reasoning, augmenting SBS with self-evaluation substantially curbs error propagation, yields higher few-shot accuracy (e.g., +6.34% to +9.56% absolute on arithmetic and strategy reasoning benchmarks), and improves calibration as measured by AUC of final answer correctness (Xie et al., 2023).
• For protein engineering, the choice of sampler (stochastic beam vs. Gibbs or denoising) is at least as consequential as model choice, with SBS obtaining top hit rates when coupled to Pareto or scalarized objective guidance (McCarter et al., 11 Mar 2026).

A summary table contextualizing SBS variants:

SBS Variant	Key Mechanism	Notable Empirical Domain
Gumbel-Top-$k$ SBS	Gumbel perturbation, SWOR	NMT, BLEU/entropy estimation
Conditional Poisson SBS	CP-SWOR, DP normalization	NMT, low-variance estimation
Beam-weighted Consistency	Prob.-weighted beams	LLM uncertainty quantification
Guided SBS (self-eval, Pareto)	Multi-objective/stepwise eval	Reasoning, protein design

7. Limitations, Open Questions, and Future Directions

Current sequence-centric SBS techniques assume white-box access to model likelihoods; black-box adaptation demands empirical probability approximations. Most experimental validations have focused on short-form generation (QA, constrained sequence design), with less clarity for long-form or unconstrained scenarios. Inclusion probabilities and estimator correction factors remain challenging to compute in large output spaces, despite DP-based tractability at moderate beam sizes.

A plausible implication is that further refinement (e.g., efficient global SWOR for massive candidate spaces, robust black-box adaptivity, hierarchical SBS) will extend the impact of sequence-centric stochastic beam search across a broader spectrum of structured generative modeling tasks.