ShredBench - 16 Fragments: leaderboard
Document restoration on ShredBench (2026): news text in two languages, source code and tables are rendered, shredded into 16 irregular, physically deformed fragments and scattered, and the model restores the content from the fragment images. Normalized edit distance to the original from 0 to 1; lower is better.
Metric: Normalized edit distance (0-1) between the restored and the original document content, aggregated over all ShredBench categories (English and Chinese news, C++, Java and Python code, tables; 756 documents) when each page is shredded into 16 irregular Voronoi fragments rendered with 3D paper deformation and scattered; the model restores the content from the fragment images; temperature 0; lower is better. Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 0.41 | #146 |
| 2 | Gemini 3 Flash | 0.44 | #180 |
| 3 | Qwen-VL-Plus | 0.65 | #5893 |
| 4 | Qwen-VL-Flash | 0.65 | |
| 5 | GLM-4.6V | 0.71 | #721 |
| 6 | InternVL3.5-38B | 0.76 | #1026 |
| 7 | ShredBench Mistral3-Reas-14B (checkpoint unspecified) | 0.77 | |
| 8 | InternVL3.5-8B | 0.78 | #1618 |
| 9 | InternVL3.5-14B | 0.78 | #1519 |
| 10 | ShredBench Mistral3-Reas-8B (checkpoint unspecified) | 0.79 | |
| 11 | GPT-5.1 | 0.82 | #248 |
| 12 | GPT-5 Mini | 0.84 | #373 |
| 13 | DeepSeek-OCR | 0.87 | |
| 14 | HunyuanOCR | 0.89 |
No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=shredbench-16-fragments · How It Works · Data refreshed daily, snapshot 2026-11-22.