Research map / Large language models / Language, safety and interpretability
On Stable Long-Form Generation: Benchmarking and Mitigating Length Volatility
Zhitao He, Haolin Yang, Rui Min, Zeyu Qin, Yi R. Fung
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply GLoBo logits boosting decoding | Qwen2.5-7B · 100-section generation | LVC (%): 17.0% -> 14.02% | Improved |
| apply GLoBo logits boosting decoding | Qwen2.5-7B · 100-section generation | MLA (%): 2.2% -> 78.25% | Improved |
| apply GLoBo logits boosting decoding | Qwen2.5-7B · 100-section generation | SCA (%): 99.8% -> 100% | Improved |
| apply GLoBo logits boosting decoding | Qwen2.5-7B · 100-section generation | UCA (%): 86.7% -> 86.7% | No clear effect |
| use GLoBo instead of repetition penalty decoding | Qwen2.5-7B · 100-section generation | LVC (%): 18.6% -> 14.02% | Improved |
| use GLoBo instead of lookahead decoding | Qwen2.5-7B · 100-section generation | LVC (%): 9.3% -> 14.02% | Improved |
| add repetition penalty decoding | Qwen2.5-7B · 100-section generation | UCA (%): 86.7% -> 76.7% | Hurt |
| add length constraint decoding | Qwen2.5-7B · 100-section generation | LVC (%): 17.0% -> 28.65% | Hurt |
| add entropy-stopping decoding | Qwen2.5-7B · 100-section generation | LVC (%): 17.0% -> 26.4% | Hurt |
| use long-text post-training (LongWriter-8B) vs Llama3.1-8B | Llama3.1-8B · 100-section generation | UCA (%): 82.0% -> 66.7% | Hurt |
| apply GLoBo to base Qwen-2.5-7B-Instruct | Qwen2.5-7B · WritingBench R3 | score: 7.4 -> 8.6 | Improved |
| apply GLoBo throughput cost | Qwen2.5-7B · throughput | tokens/s: 20.4 -> 18.2 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LongWriter-8B — LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs (ICLR 2025) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- LongGenBench: Benchmarking Long-Form Generation in Long Context LLMs (ICLR 2025 · similarity 0.91)
- LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs (ICLR 2025 · similarity 0.90)
- Precise Information Control in Long-Form Text Generation (NeurIPS 2025 · similarity 0.88)
- Graph-based Uncertainty Metrics for Long-form Language Model Generations (NeurIPS 2024 · similarity 0.88)
- Long-Short Alignment for Effective Long-Context Modeling in LLMs (ICML 2025 · similarity 0.88)
- Hansel: Output Length Controlling Framework for Large Language Models (AAAI 2025 · similarity 0.88)
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.
