Bounded length reward for sequence scoring
Mathematical Methods AIMech:BoundedLengthReward DRAFT
Aliases: Bounded word reward
A sequence scoring rule that adds a fixed nonnegative reward per generated position up to a finite length cap, giving score S(y)+r·min(l,|y|) instead of rewarding every additional position without bound.
Identity Notes & Concept Scope
The cap limits how much reward can accumulate; it is not itself a maximum allowed output length. This is the bounded reward in Huang et al. 2017, called bounded word-reward by Yang et al. 2018. It differs from division by output length, an uncapped word reward and a reward rate recomputed from beam statistics. 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
Add r·min(l,n) to cumulative log probability S(y), where n is the hypothesis length and r≥0. The 2017 method estimates l from a development-set target/source length ratio times the input length. For a current prefix at depth i, all remaining reward is bounded by r·max(l−i,0). Its total completion-score bound simplifies to S(prefix)+r·l. A beam decoder may stop once the maximum such active-prefix bound is no greater than the best completed adjusted score.
Scope
#later-name-and-scope
The 2018 paper calls the fixed-rate capped rule bounded word-reward and distinguishes it from its predicted-length and adaptive-reward variants.
⚠️ Curated Limitations & Negative Results
- The score is a decoding preference, not a probability distribution or an intrinsic translation-quality measure.
- The reward rate and length-cap estimate require specification. A development-set average length ratio need not transfer to every sentence or domain.
- The stopping bound assumes non-increasing log probability, nonnegative fixed reward and a fixed finite cap for the input; its beam-relative guarantee does not undo finite-width pruning.
- Reward saturation does not prohibit outputs longer than the cap. Tokenization, length counting and EOS handling must be specified in an implementation.
- No numerical BLEU improvement, universal quality advantage or current implementation behavior 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
|
Liang Huang, Kai Zhao, Mingbo Ma
|
primary paper
full_text
|
ID: doi:10.18653/v1/D17-1227
sha256:7b9ed5e8c842…
|
later-name-and-scope
|
Yilin Yang, Liang Huang, Mingbo Ma
|
primary paper
full_text
|
ID: doi:10.18653/v1/D18-1342
sha256:019ba14d74ad…
|
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/BoundedLengthReward.yaml