HiFlow: Hierarchical Feedback-Driven Optimization for Constrained Long-Form Text Generation

Yifan Zhu; Guanting Chen; Bing Wei; Haoran Luo

cs.CLMar 5, 2026

HiFlow: Hierarchical Feedback-Driven Optimization for Constrained Long-Form Text Generation

Authors: Yifan Zhu, Guanting Chen, Bing Wei, Haoran Luo

Organizations: Beijing University of Posts and Telecommunications · Hainan University · Nanyang Technological University

Abstract

Large language models perform well in short text generation but still struggle with long text generation, particularly under complex constraints. Such tasks involve multiple tightly coupled objectives, including global structural consistency, local semantic coherence, and constraint feasibility, forming a challenging constrained optimization problem. Existing approaches mainly rely on static planning or offline supervision, limiting effective coordination between global and local objectives during generation. To address these challenges, we propose HiFlow, a hierarchical feedback-driven optimization framework for constrained long text generation. HiFlow formulates generation as a two-level optimization process, consisting of a planning layer for global structure and constraint modeling, and a generation layer for conditioned text generation. By incorporating constraint-aware plan screening and closed-loop feedback at both levels, HiFlow enables joint optimization of planning quality and generation behavior, progressively guiding the model toward high-quality, constraint-satisfying outputs. Experiments on multiple backbones confirm HiFlow's effectiveness over baseline methods.

Figures & tables

Figure 1: An example of constrained long-form text generation in HiFlow, coordinating planning, generation, and feedback.

Figure 2: Framework comparison of HiFlow, LongWriter, and CogWriter, highlighting key differences in planning, reviewing, and generation for producing coherent, high-quality long-form text.

Figure 3: Overview of the HiFlow framework: a constrained generation process guided by reward signals, ensuring high-quality outputs, task-specific coherence, and structural consistency.

Method	Text Quality					Constraint Accuracy
	Narr.	Memory	Temporal	Affect.	Avg.	Once	Range	Peri.	Avg.
Qwen2.5-0.5B-Instruct	66.44 $\pm$ 0.25	66.13 $\pm$ 0.62	64.88 $\pm$ 1.90	76.13 $\pm$ 0.01	68.39 $\pm$ 1.01	23.46 $\pm$ 1.59	16.60 $\pm$ 1.52	15.69 $\pm$ 0.54	18.58 $\pm$ 0.75
+ CogWriter	74.75 $\pm$ 0.54	75.13 $\pm$ 0.82	72.19 $\pm$ 0.47	90.38 $\pm$ 0.91	78.11 $\pm$ 0.63	26.61 $\pm$ 0.98	19.50 $\pm$ 0.30	17.70 $\pm$ 0.24	20.78 $\pm$ 0.21
+ LongWriter	74.63 $\pm$ 0.76	74.88 $\pm$ 0.39	71.69 $\pm$ 0.68	90.00 $\pm$ 0.55	77.80 $\pm$ 0.84	21.01 $\pm$ 0.30	19.29 $\pm$ 0.11	15.83 $\pm$ 0.20	18.71 $\pm$ 0.13
+ HiFlow (ours)	75.13 $\pm$ 0.42	74.88 $\pm$ 0.61	73.19 $\pm$ 0.35	90.88 $\pm$ 0.48	78.52 $\pm$ 0.27	28.01 $\pm$ 0.55	18.53 $\pm$ 0.10	15.80 $\pm$ 0.28	22.11 $\pm$ 0.31
Qwen2.5-1.5B-Instruct	70.13 $\pm$ 0.51	69.50 $\pm$ 0.92	67.13 $\pm$ 0.43	81.50 $\pm$ 0.78	72.06 $\pm$ 0.36	23.30 $\pm$ 1.43	22.22 $\pm$ 0.28	12.12 $\pm$ 0.60	19.22 $\pm$ 0.77

Table 1: Comparison of narrative-level text quality and constraint-following accuracy across different methods.

Panel of Figure 4 — Figure 4: Results across backbone models. (a) Text quality evaluation. (b) Constraint-following accuracy. (c) Ablation study on constraint components. (d–e) Constraint accuracy and text quality on Qwen2.5-7B-Instruct. (f–g) Constraint accuracy and text quality on LLaMA3.

Panel of Figure 5 — Figure 5: Comprehensive analysis of HiFlow. (a) Case study illustrating generation quality under a representative constraint prompt. (b) Multi-dimensional performance comparison. (c) Accuracy evolution across evaluation stages. (d) Efficiency–accuracy trade-off.

Panel of Figure 6 — Figure 6: Experimental analysis across backbone models. (a–d) Planning-stage and generation-stage reward convergence on LLaMA3.1-8B and Qwen2.5-7B, showing stable optimization in both stages. (e–h) Relationship between reward scores and constraint satisfaction accuracy across different backbones, illustrating a consistent positive correlation between reward signals and constraint-aware responses.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

R(\tau)=\text{constraint satisfaction}+\text{coherence score}+\text{fidelity score}

Appendix

Algorithm 1 HiFlow: Adaptive Workflow Optimization for Constrained Long-form Generation

Text Generation

Text generation studies methods for producing coherent, useful language from prompts, data, and dialogue context.

See all 422 papers

IS-CoT: Breaking the Long-form Generation Collapse via Interleaved Structural Thinking
Jun 8, 2026Zechen Sun, Yuyang Sun, Zecheng Tang +6Open-Ended Generation Long-Form Document Generation
Institute of Computer Science and Technology, Soochow University · 2Information Research Center of Military Science, PLA Academy of Military Science
On Stable Long-Form Generation: Benchmarking and Mitigating Length Volatility
May 2, 2026Zhitao He, Haolin Yang, Rui Min +2LLM Evaluation Long-Form Document Generation
Hong Kong University of Science and Technology
Building Reliable Long-Form Generation via Hallucination Rejection Sampling
Jun 2, 2026Lin Li, Georgia Channing, Suhaas M Bhat +2Long-Form Document Generation LLM Hallucination Mitigation
Department of Computer Science, University of Oxford, Oxford, UK · Department of Engineering, University of Oxford, Oxford, UK · Oxford Digital Health Labs, Nuffield Department of Women’s and Reproductive Health, University of Oxford, Oxford, UK

Jun 8, 2026J/K move · Enter open · S save

IS-CoT: Breaking the Long-form Generation Collapse via Interleaved Structural Thinking

Zechen Sun, Yuyang Sun, Zecheng Tang, Juntao Li+5

Institute of Computer Science and Technology, Soochow University · 2Information Research Center of Military Science, PLA Academy of Military Science

Open-Ended Generation Long-Form Document Generation CoT Reasoning LLM Training Text Generation

Open paper AskPDF arXiv

Generating coherent and controllable long-form content remains a persistent challenge for Large Language Models (LLMs). While reasoning-enhanced models have demonstrated success in logic-intensive domains, our evaluation reveals that they suffer from a severe length collapse in open-ended writing, where performance degrades sharply as target lengths exceed 2,000 words. We attribute this failure to the limitation of static hierarchical planning, which struggles to provide dynamic guidance over extended contexts. To bridge this gap, we introduce the Interleaved Structural Chain-of-Thought (IS-CoT) framework. Unlike external agentic workflows, IS-CoT embeds a dynamic Plan-Write-Reflect cycle into the generation process, enabling continuous strategy adaptation and global alignment without additional assistance. Based on this framework, we construct a high-quality dataset of interleaved reasoning traces via a multi-teacher pipeline and train IS-Writer-8B. Experiments demonstrate that IS-Writer-8B achieves state-of-the-art performance on challenging long-form benchmarks (e.g., +3.08 vs. DeepSeek-V3.2 on LongBench-Write), exhibiting robust length compliance and coherence competitive with significantly larger proprietary models.

Abstract

Text Generation

Explore similar work

IS-CoT: Breaking the Long-form Generation Collapse via Interleaved Structural Thinking

On Stable Long-Form Generation: Benchmarking and Mitigating Length Volatility

Building Reliable Long-Form Generation via Hallucination Rejection Sampling