A Cognitive Writing Perspective for Constrained Long-Form Text Generation

TL;DR

Constrained long-form text generation remains difficult for single-pass LLMs. The authors formalize a cognitive-writing-inspired paradigm, CogWriter, combining a Planning Agent, multiple Generation Agents, and external monitoring and dynamic reviewing to iteratively plan, generate, and revise long outputs. Empirical results on LongGenBench-16K show significant gains over baselines, with CogWriter achieving higher instruction completion accuracy and consistent long-form generation (exceeding 10,000 words) even on backbones like Qwen-2.5-14B and GPT-4o-mini, including a 22% improvement over GPT-4o in instruction completion. The approach demonstrates the value of explicit cognitive processes for LLM writing, offering a training-free pathway to more reliable, instruction-aligned long-form text suitable for professional and research applications, albeit with higher compute cost and dependency on model capabilities.

Abstract

Like humans, Large Language Models (LLMs) struggle to generate high-quality long-form text that adheres to strict requirements in a single pass. This challenge is unsurprising, as successful human writing, according to the Cognitive Writing Theory, is a complex cognitive process involving iterative planning, translating, reviewing, and monitoring. Motivated by these cognitive principles, we aim to equip LLMs with human-like cognitive writing capabilities through CogWriter, a novel training-free framework that transforms LLM constrained long-form text generation into a systematic cognitive writing paradigm. Our framework consists of two key modules: (1) a Planning Agent that performs hierarchical planning to decompose the task, and (2) multiple Generation Agents that execute these plans in parallel. The system maintains quality via continuous monitoring and reviewing mechanisms, which evaluate outputs against specified requirements and trigger necessary revisions. CogWriter demonstrates exceptional performance on LongGenBench, a benchmark for complex constrained long-form text generation. Even when using Qwen-2.5-14B as its backbone, CogWriter surpasses GPT-4o by 22% in complex instruction completion accuracy while reliably generating texts exceeding 10,000 words. We hope this cognitive science-inspired approach provides a paradigm for LLM writing advancements: \href{https://github.com/KaiyangWan/CogWriter}{CogWriter}.

Figures (6)

• Figure 1: Comparison of human cognitive writing processes and single LLM text generation paradigm.
• Figure 2: Overview of the CogWriter Framework. The framework consists of two key modules: the Planning Agent and the Generation Agents. The Planning Agent generates and refines an initial plan, guiding the structure and flow of the document. The Generation Agents collaborate to generate, revise, and finalize document segments, ensuring consistency in content and narrative coherence across the entire document.
• Figure 3: Comparison of Length Control Ability.
• Figure 4: Comparison of Instruction Type Performance.
• Figure 5: Length Control Performance Across Different Models and Task Types. (a) and (c) show performance on spatial tasks requiring 150 words per unit, while (b) and (d) present results for temporal tasks with 200-word requirements.