Evidence-First Generation: A Multi-Granular Framework for Evidence-Prioritized Autoregressive Text Generation, Content Selection, and Dynamic Reorganization
Abstract
Autoregressive language models typically generate text by predicting successive tokens conditioned on previously generated tokens and available contextual information. Although this approach supports fluent and contextually coherent language generation, it does not inherently require a system to collect and evaluate external evidence before deciding what content to express, nor does it explicitly require evidence value to determine the order in which words, sentences, or paragraphs are developed. Existing research has investigated retrieval-augmented generation, evidence-grounded text generation, content planning, sentence-level factual generation, and document-level organization. These studies establish important foundations for evidence-oriented writing systems, but they do not collectively resolve the broader design question of how evidence acquisition, content selection, generation granularity, and output ordering can be jointly controlled by a unified mechanism.
This paper proposes Evidence-First Generation (EFG), a conceptual framework that places evidence acquisition and assessment before, or in continuous interaction with, textual realization. EFG represents evidence as structured units, evaluates candidate content according to evidential support, task relevance, information value, uncertainty, redundancy, and dependency constraints, and selects an appropriate generation strategy at token, phrase, sentence, paragraph, or document level. The framework supports both ordered generation, in which selected content is rearranged into a coherent sequence, and order-preserving generation, in which content is realized according to a predefined structural plan. It also permits hybrid strategies that alternate between evidence collection, content generation, verification, and revision.
The paper formalizes the proposed framework, distinguishes it from related approaches, develops a set of testable research hypotheses, and outlines an experimental protocol for comparing EFG with conventional autoregressive generation, retrieval-augmented generation, and planning-based systems. The framework is a research proposal rather than an experimentally validated model. Its principal contribution is a unified design space for investigating evidence-prioritized generation across multiple textual granularities, with explicit treatment of evidence quality, content dependencies, ordering decisions, and generation costs.
Related
- Problem-First Writing: A Question-Centered Framework for Multi-Granular Content Selection, Generation, and Organization in Autoregressive Language Models
- Position-Wise Dynamic Competition for Autoregressive Text Generation: A Framework for Adaptive Token-Position Selection
- 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
- Granular Deposition Image Generation: A Framework for Target-Guided Visual Synthesis Through Particle Transport, Selective Deposition, and Dynamic Reorganization
BibTeX
@misc{zhang2026evidencefirst,
title = {{Evidence-First Generation: A Multi-Granular Framework for Evidence-Prioritized Autoregressive Text Generation, Content Selection, and Dynamic Reorganization}},
author = {Jincheng Zhang},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23614906},
url = {https://zenodo.org/records/23614906},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Evidence-First Generation: A Multi-Granular Framework for Evidence-Prioritized Autoregressive Text Generation, Content Selection, and Dynamic Reorganization
AU - Jincheng Zhang
PY - 2026
DA - 2026/10/24
PB - Zenodo
DO - 10.5281/zenodo.23614906
UR - https://zenodo.org/records/23614906
AB - Autoregressive language models typically generate text by predicting successive tokens conditioned on previously generated tokens and available contextual information. Although this approach supports fluent and contextually coherent language generation, it does not inherently require a system to collect and evaluate external evidence before deciding what content to express, nor does it explicitly require evidence value to determine the order in which words, sentences, or paragraphs are developed. Existing research has investigated retrieval-augmented generation, evidence-grounded text generation, content planning, sentence-level factual generation, and document-level organization. These studies establish important foundations for evidence-oriented writing systems, but they do not collectively resolve the broader design question of how evidence acquisition, content selection, generation granularity, and output ordering can be jointly controlled by a unified mechanism. This paper proposes Evidence-First Generation (EFG), a conceptual framework that places evidence acquisition and assessment before, or in continuous interaction with, textual realization. EFG represents evidence as structured units, evaluates candidate content according to evidential support, task relevance, information value, uncertainty, redundancy, and dependency constraints, and selects an appropriate generation strategy at token, phrase, sentence, paragraph, or document level. The framework supports both ordered generation, in which selected content is rearranged into a coherent sequence, and order-preserving generation, in which content is realized according to a predefined structural plan. It also permits hybrid strategies that alternate between evidence collection, content generation, verification, and revision. The paper formalizes the proposed framework, distinguishes it from related approaches, develops a set of testable research hypotheses, and outlines an experimental protocol for comparing EFG with conventional autoregressive generation, retrieval-augmented generation, and planning-based systems. The framework is a research proposal rather than an experimentally validated model. Its principal contribution is a unified design space for investigating evidence-prioritized generation across multiple textual granularities, with explicit treatment of evidence quality, content dependencies, ordering decisions, and generation costs.
ER -