Vanilla sequence beam decoding
Algorithms AIMech:VanillaSequenceBeamDecoding DRAFT
Aliases: Vanilla beam decoding
A layer-based sequence beam decoder in which completed and unfinished hypotheses compete for the same k beam slots: finished hypotheses are carried forward unchanged, unfinished hypotheses are extended, and the top-scoring candidates form the next beam.
Identity Notes & Concept Scope
This is the vanilla sequence-decoding policy contrasted with FCFS by Kasai and colleagues. It specializes generic breadth-first beam search by specifying EOS retention and competition for active slots. Finished outputs are not accumulated in a separate permanent set, and the procedure is not the score-prioritized per-length best-first decoder. No normalized identity match was found across the 442 AIO and 141 prior native records; ignored-inclusive text hits in the main checkout were prior discovery leads and incidental substrings such as builds in records, source text and dependencies, not existing concept records.
Claims & Grounded Evidence
2 Claims
Mechanism
#search-procedure
At each sequence length, construct a fresh candidate pool from carried-forward EOS-complete hypotheses and one-token extensions of unfinished beam members. Select up to k highest-scoring candidates for the next beam. The 2024 description stops when all retained hypotheses are complete, or at the maximum length. The 2020 standard-beam pseudocode explicitly resets the candidate pool at each layer and carries finished hypotheses unchanged.
Mechanism
#fresh-layer-pool
The standard sequence-beam algorithm initializes an empty candidate pool for each time step, carries EOS-complete hypotheses unchanged, extends unfinished hypotheses and selects the next top-k beam. This supplies an explicit layer-pool scope for the later vanilla-policy description.
⚠️ Curated Limitations & Negative Results
- A completed hypothesis can later fall off the beam because it competes with unfinished candidates. Its discovery alone does not reserve an output slot.
- Finite beam pruning is not exact global MAP decoding. Different tie, length-score and maximum-length policies can change outputs.
- At the length limit, an implementation must specify what to return if no retained hypothesis is EOS-complete; the simplified figures do not provide a universal failure policy.
- Figure 2 in the 2024 paper omits initialization/reset of its temporary H pool. The fresh-per-layer interpretation is supported by the layer-based prose and the explicit reset in Meister et al. 2020 Algorithm 1; the record does not certify the 2024 figure as executable code.
- The name identifies the source-defined policy, not a claim that all libraries or all uses of beam search implement it.
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 these sequence-decoding policies for natural-language generation, including translation and summarization. The exact AIO CLASS describes the language-processing task. This is an algorithm-to-application-domain relationship, not equivalence or an is-a assertion; the decoder is not a trained language model and can also be used outside NLP. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Dragomir Radev, Yejin Choi, Noah A. Smith
|
primary paper
full_text
|
sha256:7d090ee35511…
|
fresh-layer-pool
|
Clara Meister, Tim Vieira, Ryan Cotterell
|
primary paper
full_text
|
ID: doi:10.1162/tacl_a_00346
sha256:46728395bd4f…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T06:26:52Z
Updated: 2026-10-11T06:29:45Z
Source File:
data/curated/VanillaSequenceBeamDecoding.yaml