Absolute threshold pruning for sequence beams
Algorithms AIMech:AbsoluteThresholdBeamPruning DRAFT
A sequence-beam pruning procedure that first applies the beam-size cap, then discards candidates whose score is at least a specified positive margin below the best score in that capped candidate list.
Identity Notes & Concept Scope
This is Freitag and Al-Onaizan’s absolute threshold pruning rule. It uses an additive score difference, unlike their relative-score and last-token relative thresholds. It is an additional local pruning step, not a stopping certificate or a method that automatically widens the beam. 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
First retain candidates up to the current beam width. Let M be the maximum score in that candidate list and let a>0 be the absolute pruning margin. Discard a candidate when its score is at most M−a, including equality as printed in Equation (2). Continue decoding with the survivors; applying pruning after the cap can reduce the active count without refilling from previously discarded candidates.
Limitation
#search-error-scope
Additional pruning can change outputs by discarding a better-scoring hypothesis. The paper’s reported quality preservation and speed improvements are empirical outcomes in specific language-pair setups, not exact-search guarantees.
⚠️ Curated Limitations & Negative Results
- The extra pruning can remove a prefix leading to the best eventual output. The paper explicitly reports search errors even where aggregate task metrics remain similar.
- The numerical margin depends on score scale, log base and model calibration. It is not a relative probability threshold.
- A positive margin is required to retain the best finite-scoring item under the printed non-strict inequality; a zero margin would discard even maximum-score candidates.
- The procedure is layered onto a decoder with its own EOS and width policy. It does not by itself specify the entire decoder or guarantee improved speed or quality.
- The separate relative-pruning equations have a score-scale ambiguity under negative log scores. No correction to those distinct variants is inferred from this additive rule.
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
|
Markus Freitag, Yaser Al-Onaizan
|
primary paper
full_text
|
ID: doi:10.18653/v1/W17-3207
sha256:7d157efa8f64…
|
search-error-scope
|
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/AbsoluteThresholdBeamPruning.yaml