Position-Wise Dynamic Competition for Autoregressive Text Generation: A Framework for Adaptive Token-Position Selection
Abstract
Autoregressive language models conventionally generate text through a sequential process in which the next token is predicted conditional on previously generated tokens. Although this formulation has achieved substantial success, its standard left-to-right decoding procedure restricts the order in which output positions are resolved. Existing insertion-based and non-monotonic generation methods relax this restriction, but the relationship between dynamic competition among unresolved token positions and the order in which these positions are finalized remains an important design question.
This paper proposes Position-Wise Dynamic Competition (PWDC), a framework for investigating adaptive competition among unresolved positions during text generation. Instead of treating the next generation step as an unconditional commitment to the next left-to-right position, PWDC maintains a set of active positions, evaluates their current readiness and expected contribution, and dynamically allocates generation opportunities through a competition mechanism. Candidate tokens are evaluated locally at each position, while positions compete globally for the next update. The proposed framework distinguishes candidate-token selection from position scheduling and final sequence ordering. Its components include position-level utility estimation, competition-aware scheduling, confidence-sensitive commitment, dependency-aware consistency checking, and adaptive reconsideration of unresolved positions.
A unified mathematical formulation is introduced, together with a training objective, an inference procedure, a computational analysis, and an experimental protocol. The proposed method is intended to test whether adaptive position scheduling can improve generation quality, constraint satisfaction, or computational efficiency under appropriate conditions. Because insertion-based decoding, arbitrary-order generation, and token-order learning have already been investigated, the contribution is framed as a testable integration of dynamic position competition and adaptive commitment rather than a claim that non-left-to-right generation is new. Empirical evaluation is required to establish its effectiveness.
Related
- Dynamic Sentence-Position Competition for Text Generation: A Framework for Candidate Competition, Joint Ordering, and Global Coherence Optimization
- Dynamic Competition Among Paragraph Positions: A Position-Aware Framework for Adaptive Long-Form Text Generation
- Dynamic Competition for Sentence Positions in Text Generation: A Candidate–Position Matching Framework for Adaptive Sentence Generation and Global Text Organization
- Evidence-First Generation: A Multi-Granular Framework for Evidence-Prioritized Autoregressive Text Generation, Content Selection, and Dynamic Reorganization
- Problem-First Writing: A Question-Centered Framework for Multi-Granular Content Selection, Generation, and Organization in Autoregressive Language Models
BibTeX
@misc{zhang2026positionwise,
title = {{Position-Wise Dynamic Competition for Autoregressive Text Generation: A Framework for Adaptive Token-Position Selection}},
author = {Jincheng Zhang},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23614685},
url = {https://zenodo.org/records/23614685},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Position-Wise Dynamic Competition for Autoregressive Text Generation: A Framework for Adaptive Token-Position Selection
AU - Jincheng Zhang
PY - 2026
DA - 2026/10/24
PB - Zenodo
DO - 10.5281/zenodo.23614685
UR - https://zenodo.org/records/23614685
AB - Autoregressive language models conventionally generate text through a sequential process in which the next token is predicted conditional on previously generated tokens. Although this formulation has achieved substantial success, its standard left-to-right decoding procedure restricts the order in which output positions are resolved. Existing insertion-based and non-monotonic generation methods relax this restriction, but the relationship between dynamic competition among unresolved token positions and the order in which these positions are finalized remains an important design question. This paper proposes Position-Wise Dynamic Competition (PWDC), a framework for investigating adaptive competition among unresolved positions during text generation. Instead of treating the next generation step as an unconditional commitment to the next left-to-right position, PWDC maintains a set of active positions, evaluates their current readiness and expected contribution, and dynamically allocates generation opportunities through a competition mechanism. Candidate tokens are evaluated locally at each position, while positions compete globally for the next update. The proposed framework distinguishes candidate-token selection from position scheduling and final sequence ordering. Its components include position-level utility estimation, competition-aware scheduling, confidence-sensitive commitment, dependency-aware consistency checking, and adaptive reconsideration of unresolved positions. A unified mathematical formulation is introduced, together with a training objective, an inference procedure, a computational analysis, and an experimental protocol. The proposed method is intended to test whether adaptive position scheduling can improve generation quality, constraint satisfaction, or computational efficiency under appropriate conditions. Because insertion-based decoding, arbitrary-order generation, and token-order learning have already been investigated, the contribution is framed as a testable integration of dynamic position competition and adaptive commitment rather than a claim that non-left-to-right generation is new. Empirical evaluation is required to establish its effectiveness.
ER -