Dynamic Sentence-Position Competition for Text Generation: A Framework for Candidate Competition, Joint Ordering, and Global Coherence Optimization
Abstract
Autoregressive language models generate text through sequential token prediction, whereas sentence-level content planning and sentence ordering have been studied as complementary mechanisms for improving document organization. However, a unified framework that explicitly models repeated competition among mutable sentence candidates, dynamically assigns candidates to structural positions, and jointly optimizes sentence selection and document-level coherence remains worth investigating. This paper proposes Dynamic Sentence-Position Competition (DSPC), a framework for sentence-oriented text generation in which candidate sentences compete for retention and placement under multiple content and structural criteria. Unlike a pipeline that generates a fixed set of sentences and subsequently sorts them, DSPC maintains a mutable candidate pool and an evolving assignment between candidates and sentence positions. The framework combines task relevance, information gain, semantic contribution, redundancy, dependency compatibility, transition coherence, and generation cost. A position-aware competition mechanism updates candidate priorities, while a joint selection-and-ordering objective coordinates content value and global textual structure. An iterative feedback mechanism allows the provisional document to influence subsequent candidate revision and competition. We formulate the framework mathematically, describe a practical implementation compatible with autoregressive language models, and propose an experimental protocol comparing it with conventional autoregressive generation, generate-then-rerank pipelines, sentence-planning methods, and iterative refinement approaches. The framework is a research proposal rather than an empirically validated model. Its principal hypothesis is that explicit competition over both sentence retention and structural position may improve the controllability and organization of generated texts under suitable evaluation conditions.
Related
- Dynamic Competition for Sentence Positions in Text Generation: A Candidate–Position Matching Framework for Adaptive Sentence Generation and Global Text Organization
- Territory-Competition Image Generation: A Dynamic Framework for Adaptive Spatial Allocation, Region Evolution, and Reward-Driven Visual Synthesis
- Pixel Placement Generation: A Puzzle-Inspired Framework for Image Synthesis Through Sequential Placement, Spatial Competition, and Iterative Rearrangement
- Interaction-and-Competition-Driven 3D Generation: A Multi-Agent Framework for Emergent Volumetric Structure Synthesis
- Foam-Based Image Generation: A Dynamic Framework for Bubble Formation, Spatial Competition, and Boundary-Guided Visual Synthesis
BibTeX
@misc{zhang2026dynamic,
title = {{Dynamic Sentence-Position Competition for Text Generation: A Framework for Candidate Competition, Joint Ordering, and Global Coherence Optimization}},
author = {Jincheng Zhang},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23614764},
url = {https://zenodo.org/records/23614764},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Dynamic Sentence-Position Competition for Text Generation: A Framework for Candidate Competition, Joint Ordering, and Global Coherence Optimization
AU - Jincheng Zhang
PY - 2026
DA - 2026/10/24
PB - Zenodo
DO - 10.5281/zenodo.23614764
UR - https://zenodo.org/records/23614764
AB - Autoregressive language models generate text through sequential token prediction, whereas sentence-level content planning and sentence ordering have been studied as complementary mechanisms for improving document organization. However, a unified framework that explicitly models repeated competition among mutable sentence candidates, dynamically assigns candidates to structural positions, and jointly optimizes sentence selection and document-level coherence remains worth investigating. This paper proposes Dynamic Sentence-Position Competition (DSPC), a framework for sentence-oriented text generation in which candidate sentences compete for retention and placement under multiple content and structural criteria. Unlike a pipeline that generates a fixed set of sentences and subsequently sorts them, DSPC maintains a mutable candidate pool and an evolving assignment between candidates and sentence positions. The framework combines task relevance, information gain, semantic contribution, redundancy, dependency compatibility, transition coherence, and generation cost. A position-aware competition mechanism updates candidate priorities, while a joint selection-and-ordering objective coordinates content value and global textual structure. An iterative feedback mechanism allows the provisional document to influence subsequent candidate revision and competition. We formulate the framework mathematically, describe a practical implementation compatible with autoregressive language models, and propose an experimental protocol comparing it with conventional autoregressive generation, generate-then-rerank pipelines, sentence-planning methods, and iterative refinement approaches. The framework is a research proposal rather than an empirically validated model. Its principal hypothesis is that explicit competition over both sentence retention and structural position may improve the controllability and organization of generated texts under suitable evaluation conditions.
ER -