Home / Curated / Algorithms / Maximum candidates per node in sequence beam pruning

Maximum candidates per node in sequence beam pruning

Algorithms AIMech:PerNodeCandidateBeamPruning DRAFT

A sequence-beam pruning procedure that limits how many retained candidate extensions may come from the same predecessor hypothesis, discarding a candidate when the allowed number of better-scoring siblings is already retained.

Identity Notes & Concept Scope This is the Maximum Candidates per Node rule of Freitag and Al-Onaizan. Same history means the common predecessor prefix, not arbitrary semantic similarity between full sequences. It differs from an additive diversity penalty, grouping beams into separate searches or imposing a global frontier capacity. No normalized identity match was found across the 442 AIO and 146 prior native records; both checkout text searches included ignored and hidden files and returned no matches for the selected exact IDs, full labels, aliases and source titles.

Claims & Grounded Evidence

2 Claims

Mechanism #method

Apply the current beam-size cap, then keep at most m candidate continuations for each predecessor prefix by score. A candidate is discarded once m better-scoring candidates with that same predecessor are in the retained list. This is extra pruning after the initial top-beam selection; it does not specify refilling with candidates outside that initial list.

supports Source: Beam Search Strategies for Neural Machine Translation · Locator: Section 4, introductory paragraph and Maximum Candidates per Node, printed p. 57 / PDF p. 2

Limitation #diversity-limit

The source introduced a per-history candidate cap to diversify predecessor coverage, but its conclusion states that adding diversity did not improve translation quality in the evaluated settings.

supports Source: Beam Search Strategies for Neural Machine Translation · Locator: Introduction and Section 6, printed pp. 56 and 58 / PDF pp. 1 and 3

⚠️ Curated Limitations & Negative Results

A positive integer per-parent cap and deterministic tie handling must be specified. The source’s prose refers to better-scoring candidates without a complete tie policy.
Limiting sibling extensions can remove the prefix of the best eventual sequence; this is not an exact or complete search guarantee.
More varied parent prefixes do not guarantee semantic diversity or improved translation quality. The paper’s conclusion reports no translation-quality improvement from this diversity intervention.
The inherited decoder determines EOS processing, generation length and changing active-beam width. The pruning rule alone is not a complete decoder implementation.
No specific current library implementation, numeric quality gain or universal runtime benefit is asserted.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:NaturalLanguageProcessing` Natural Language Processing	related_match	The primary paper defines this sequence scoring, stopping or pruning method for natural-language generation, with neural machine translation as its central setting. The exact AIO CLASS describes the language-processing task. This is an method-to-application-domain relationship, not equivalence or an is-a assertion; the method is not a trained language-model instance and may also apply outside NLP.

Evidence Sources & Primary Literature

2 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`method-source`	Beam Search Strategies for Neural Machine Translation Markus Freitag, Yaser Al-Onaizan	primary paper full_text	ID: doi:10.18653/v1/W17-3207 sha256:7d157efa8f64…
`diversity-limit`	Beam Search Strategies for Neural Machine Translation Markus Freitag, Yaser Al-Onaizan	primary paper full_text	ID: doi:10.18653/v1/W17-3207 sha256:7d157efa8f64…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-11T06:36:44Z

Updated: 2026-10-11T06:37:11Z

Source File: data/curated/PerNodeCandidateBeamPruning.yaml