Dynamic Competition Among Paragraph Positions: A Position-Aware Framework for Adaptive Long-Form Text Generation
Abstract
Autoregressive language models typically generate text from left to right, making the current generation position an implicit constraint on when individual parts of a document can be developed. Existing approaches to long-form text generation have explored content planning, sentence-level positioning, iterative refinement, and non-left-to-right insertion. However, the question of whether paragraph positions can be modeled as competing generation opportunities whose priorities change dynamically throughout document construction remains worthy of explicit investigation. This paper proposes Dynamic Competition Among Paragraph Positions (DCPP), a framework in which a document is represented as a set of partially specified paragraph slots. At each generation step, a controller evaluates the current utility of available positions according to task relevance, information gain, structural importance, dependency readiness, redundancy, uncertainty, and estimated generation cost. A competition mechanism allocates generation opportunities among these positions, while dependency constraints, positional stability, and global coherence objectives regulate switching and placement. The framework supports both fixed-slot and expandable-document settings and can operate with an autoregressive language model as its paragraph generator. We formalize the position-selection problem, introduce a cost-aware utility function, specify a training objective and inference algorithm, and propose an experimental protocol comparing DCPP with left-to-right generation, fixed content planning, and iterative refinement baselines. The framework is intended to test whether adaptive competition over paragraph positions can improve long-document coherence, content coverage, redundancy control, and generation efficiency. Its effectiveness and novelty relative to existing dynamic content-planning methods remain empirical questions.
Related
- Dynamic Competition for Sentence Positions in Text Generation: A Candidate–Position Matching Framework for Adaptive Sentence Generation and Global Text Organization
- Dynamic Competitive Paragraph Generation and Ordering: A Position-Aware Framework for Candidate Selection, Global Reordering, and Coherent Text Integration
- Dynamic Sentence-Position Competition for Text Generation: A Framework for Candidate Competition, Joint Ordering, and Global Coherence Optimization
- Position-Wise Dynamic Competition for Autoregressive Text Generation: A Framework for Adaptive Token-Position Selection
- Territory-Competition Image Generation: A Dynamic Framework for Adaptive Spatial Allocation, Region Evolution, and Reward-Driven Visual Synthesis
BibTeX
@misc{zhang2026dynamic,
title = {{Dynamic Competition Among Paragraph Positions: A Position-Aware Framework for Adaptive Long-Form Text Generation}},
author = {Jincheng Zhang},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23614720},
url = {https://zenodo.org/records/23614720},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Dynamic Competition Among Paragraph Positions: A Position-Aware Framework for Adaptive Long-Form Text Generation
AU - Jincheng Zhang
PY - 2026
DA - 2026/10/24
PB - Zenodo
DO - 10.5281/zenodo.23614720
UR - https://zenodo.org/records/23614720
AB - Autoregressive language models typically generate text from left to right, making the current generation position an implicit constraint on when individual parts of a document can be developed. Existing approaches to long-form text generation have explored content planning, sentence-level positioning, iterative refinement, and non-left-to-right insertion. However, the question of whether paragraph positions can be modeled as competing generation opportunities whose priorities change dynamically throughout document construction remains worthy of explicit investigation. This paper proposes Dynamic Competition Among Paragraph Positions (DCPP), a framework in which a document is represented as a set of partially specified paragraph slots. At each generation step, a controller evaluates the current utility of available positions according to task relevance, information gain, structural importance, dependency readiness, redundancy, uncertainty, and estimated generation cost. A competition mechanism allocates generation opportunities among these positions, while dependency constraints, positional stability, and global coherence objectives regulate switching and placement. The framework supports both fixed-slot and expandable-document settings and can operate with an autoregressive language model as its paragraph generator. We formalize the position-selection problem, introduce a cost-aware utility function, specify a training objective and inference algorithm, and propose an experimental protocol comparing DCPP with left-to-right generation, fixed content planning, and iterative refinement baselines. The framework is intended to test whether adaptive competition over paragraph positions can improve long-document coherence, content coverage, redundancy control, and generation efficiency. Its effectiveness and novelty relative to existing dynamic content-planning methods remain empirical questions.
ER -