Problem-First Writing: A Question-Centered Framework for Multi-Granular Content Selection, Generation, and Organization in Autoregressive Language Models
Abstract
Autoregressive language models typically generate text through sequential token prediction, in which each new token is conditioned on the preceding context. Although this mechanism can produce coherent text, it does not inherently require the generation process to begin by explicitly identifying the questions that a document must answer. Human writers, by contrast, may begin by identifying a problem, decomposing it into subquestions, selecting relevant content, and deciding how individual pieces of information should be organized. This paper proposes Problem-First Writing (PFW), a question-centered framework for natural language generation that makes problem identification an explicit control principle. PFW represents a writing task as a set of answerable questions and associates candidate content units with the questions they address. The framework supports multiple granularities, including words, phrases, sentences, paragraphs, and combinations of these units. Candidate units may be generated sequentially, generated independently, selected from existing text, or produced through a hybrid process. An optional organization module orders and connects the selected units according to answer dependencies, discourse relations, redundancy, and global coherence. A further evaluation stage assesses whether the resulting document adequately addresses its original questions. This paper formalizes the framework, distinguishes several implementation architectures, proposes objective functions and evaluation metrics, and develops an experimental protocol for comparing PFW with conventional autoregressive generation and existing planning-based methods. The framework is motivated by prior work on question-answer blueprints, content planning, and sentence ordering; consequently, its contribution is presented as a unifying research proposal rather than an established claim of absolute novelty. The central research question is whether explicitly controlling generation through problem-centered content selection can improve answer completeness, relevance, organization, and efficiency without unnecessarily sacrificing fluency or flexibility.
Related
- Evidence-First Generation: A Multi-Granular Framework for Evidence-Prioritized Autoregressive Text Generation, Content Selection, and Dynamic Reorganization
- Position-Wise Dynamic Competition for Autoregressive Text Generation: A Framework for Adaptive Token-Position Selection
- The right measure for the right question: A framework for domain‐matched patient‐reported outcome measure selection in knee arthroplasty research
- Meta Non-Autoregressive Models: A Framework for Generating, Selecting, and Adapting Non-Autoregressive Generation Mechanisms
- Structure-First Text Generation: A Unified Framework for Hierarchical Planning, Content Realization, Ordering, and Composition in Autoregressive Language Models
BibTeX
@misc{zhang2026problemfirst,
title = {{Problem-First Writing: A Question-Centered Framework for Multi-Granular Content Selection, Generation, and Organization in Autoregressive Language Models}},
author = {Jincheng Zhang},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23614932},
url = {https://zenodo.org/records/23614932},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Problem-First Writing: A Question-Centered Framework for Multi-Granular Content Selection, Generation, and Organization in Autoregressive Language Models
AU - Jincheng Zhang
PY - 2026
DA - 2026/10/24
PB - Zenodo
DO - 10.5281/zenodo.23614932
UR - https://zenodo.org/records/23614932
AB - Autoregressive language models typically generate text through sequential token prediction, in which each new token is conditioned on the preceding context. Although this mechanism can produce coherent text, it does not inherently require the generation process to begin by explicitly identifying the questions that a document must answer. Human writers, by contrast, may begin by identifying a problem, decomposing it into subquestions, selecting relevant content, and deciding how individual pieces of information should be organized. This paper proposes Problem-First Writing (PFW), a question-centered framework for natural language generation that makes problem identification an explicit control principle. PFW represents a writing task as a set of answerable questions and associates candidate content units with the questions they address. The framework supports multiple granularities, including words, phrases, sentences, paragraphs, and combinations of these units. Candidate units may be generated sequentially, generated independently, selected from existing text, or produced through a hybrid process. An optional organization module orders and connects the selected units according to answer dependencies, discourse relations, redundancy, and global coherence. A further evaluation stage assesses whether the resulting document adequately addresses its original questions. This paper formalizes the framework, distinguishes several implementation architectures, proposes objective functions and evaluation metrics, and develops an experimental protocol for comparing PFW with conventional autoregressive generation and existing planning-based methods. The framework is motivated by prior work on question-answer blueprints, content planning, and sentence ordering; consequently, its contribution is presented as a unifying research proposal rather than an established claim of absolute novelty. The central research question is whether explicitly controlling generation through problem-centered content selection can improve answer completeness, relevance, organization, and efficiency without unnecessarily sacrificing fluency or flexibility.
ER -