VOLTBench - Mean Length Accuracy (50 Sections): leaderboard

Length adherence on VOLTBench (2026), a long-form generation benchmark of chapter-based story and code instructions, at the 50-section target: five generations are sampled per instruction and their mean length is scored as 100 x max(0, 1 - |mean - target| / target). Percent; higher is better.

Metric: Mean length accuracy (%): 100 x max(0, 1 - |mean output length - target length| / target length) over the five sampled generations per instruction, English simple-difficulty instructions, Story (unstructured) and Code Function (structured) chapter-based tasks, 50-section target; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-4o Mini55.3#1667
2LongWriter-8B51.5
3DeepSeek V319#729
4Falcon3-Mamba-7B-Instruct15.2
5Qwen 2.5 7B Instruct13.7#2723
6DeepSeek R18.3#482
7Qwen 2.5 1.5B Instruct6.4#7470
8Llama 3.1 8B Instruct2.8#4600
9Claude 3.5 Sonnet1.6#752

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=voltbench-mean-length-accuracy-50-sections · How It Works · Data refreshed daily, snapshot 2026-11-21.