Eliminating Agentic Workflow for Introduction Generation with Parametric Stage Tokens

TL;DR

STIG eliminates external agentic workflows by embedding stage semantics into parametric tokens, enabling end-to-end Introduction generation in a single inference. Trained on thousands of ACL papers and using eight stage-token pairs to govern four subsections, STIG demonstrates superior semantic fidelity, structural rationality, and content coverage while delivering up to $3.3\times$ efficiency gains over agentic baselines. Quantitative and human evaluations show STIG outperforms traditional and contemporary baselines across multiple metrics, validating its approach to structure-aware, single-pass generation. The work suggests a practical path toward scalable, coherent scientific writing by internalizing phased reasoning within the LLM parameters.

Abstract

In recent years, using predefined agentic workflows to guide large language models (LLMs) for literature classification and review has become a research focus. However, writing research introductions is more challenging. It requires rigorous logic, coherent structure, and abstract summarization. Existing workflows often suffer from long reasoning chains, error accumulation, and reduced textual coherence. To address these limitations, we propose eliminating external agentic workflows. Instead, we directly parameterize their logical structure into the LLM. This allows the generation of a complete introduction in a single inference. To this end, we introduce the Stage Token for Introduction Generation (STIG). STIG converts the multiple stages of the original workflow into explicit stage signals. These signals guide the model to follow different logical roles and functions during generation. Through instruction tuning, the model learns the mapping between stage tokens and text functions. It also learns the logical order and transition patterns between stages, encoding this knowledge into the model parameters. Experimental results show that STIG can generate multi-stage text in a single inference. It does not require explicit workflow calls. STIG outperforms traditional agentic workflows and other baselines on metrics of semantic similarity and sentence-level structural rationality. The code is provided in the Supplementary Materials.

Figures (4)

• Figure 1: Comparison of Agentic Workflow and STIG in writing. Agentic workflows rely on multi-turn interactions. They complete writing step by step, including outline generation, subsection drafting, integration, and refinement. In contrast, STIG generates the content in a single inference. The final output is obtained by parsing the generated text according to stage tokens.
• Figure 2: Overview of the STIG Model Architecture. STIG integrates eight stage-token pairs that guide the generation of four subsections, across outline and content phases. We highlight the flow from core input materials to structured output, emphasizing the single inference mechanism and the role of parametric stage tokens in enhancing logical coherence. The top-right corner illustrates the stage writing finetuning procedure, which constitutes a component of our ablation study in section \ref{['subsection:ablation']}.
• Figure 3: The introduction generated by AutoSurvey, with each sentence annotated to correspond to one of the four subsections. The introduction generated by AutoSurvey exhibits a highly irrational structure, with excessive and repetitive emphasis on the method. Detailed introduction is shown in \ref{['app:auto_2']}.
• Figure 4: Statistics on token consumption and effectively generated tokens. Our method demonstrates the highest token usage efficiency.