Research map / Large language models / Alignment and preferences
LoopWeaver: Weaving Feedback Loops into Hierarchical Generation under Constraints
Yifan Zhu, Guanting Chen, Bing Wei, Haoran Luo
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LoopWeaver, a feedback-loop approach to hierarchical LLM generation for long text under constraints, jointly optimizing global structure, local coherence, and constraint satisfaction.
Problem
Constrained long-text generation with LLMs requires jointly balancing global structure, local coherence, and constraint satisfaction, but existing static or loosely coupled pipelines struggle with cross-stage collaborative optimization.
Approach
Couples hierarchical generation stages through feedback loops so that stages collaboratively optimize rather than running as a static pipeline (abstract is truncated, so the specific mechanism is not stated).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply LoopWeaver planning + filtering + two-stage DPO | Qwen2.5-7B-Instruct · LongGenBench-style constraint accuracy | avg accuracy: 30.74 -> 48.41 | Improved |
| apply LoopWeaver | Qwen2.5-7B-Instruct · LongGenBench-style text quality | avg score: 71.34 -> 78.77 | Improved |
| apply LoopWeaver | Qwen2.5-0.5B-Instruct · LongGenBench periodic constraints | Peri. accuracy: 15.69 -> 15.80 (CogWriter 17.70) | No clear effect |
| apply LoopWeaver | Qwen2.5-0.5B-Instruct · LongGenBench affective consistency | Affect. score: 39.86 -> 37.39 | Hurt |
| apply CogWriter | Qwen2.5-0.5B-Instruct · LongGenBench text quality | avg score: 27.49 -> 25.16 | Hurt |
| apply LongWriter | Qwen2.5-7B-Instruct · LongGenBench constraint accuracy | avg accuracy: 30.74 -> 29.21 | No clear effect |
| add plan-level DPO only (ablation) | Qwen2.5-7B-Instruct · LongGenBench constraint accuracy | avg accuracy: 30.74 -> 44.20 | Improved |
| add generation-level DPO only (ablation) | Qwen2.5-7B-Instruct · LongGenBench constraint accuracy | avg accuracy: 30.74 -> 44.78 | Improved |
| combine both DPO stages and workflow (ablation) | Qwen2.5-7B-Instruct · LongGenBench constraint accuracy | avg accuracy: 44.78 -> 48.41 | Improved |
| combine both stages vs best single stage (ablation) | Qwen2.5-7B-Instruct · LongGenBench text quality | avg score: 77.23 -> 78.77 | Improved |
| vary reward threshold tau 0.7 -> 0.6 or 0.8 (ablation) | reward construction · Fixed-candidate audit | selection retained: 97.6% / 99.2% retained | No clear effect |
| feasibility filter on plans | Llama-3.1-8B · Filter audit 800 candidates | proxy precision/recall: 61.0 / 72.8 percent; false rejection 27.2% | Mixed |
Closest papers by meaning
- LongGenBench: Benchmarking Long-Form Generation in Long Context LLMs (ICLR 2025 · similarity 0.86)
- Guaranteed Generation from Large Language Models (ICLR 2025 · similarity 0.85)
- ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution (NeurIPS 2024 · similarity 0.85)
- Planning in Natural Language Improves LLM Search for Code Generation (ICLR 2025 · similarity 0.85)
- WritingBench: A Comprehensive Benchmark for Generative Writing (NeurIPS 2025 · similarity 0.85)
- Reflection-Window Decoding: Text Generation with Selective Refinement (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
