LongDocBench - Contextual Relationship Recovery: leaderboard
Recovery of the caption, note and source relationships of tables and figures (macro mean) in LongDocBench (2026), given a fixed parse and located objects. Normalized text edit similarity in percent, higher is better.
Metric: Macro similarity (%; unweighted mean of the caption, note and source relationship scores; the model reads fixed TextIn parsing output and layout context of long financial reports, textbooks and papers and links each of 2,680 benchmark-localized tables and figures to its related text; normalized edit similarity to the gold text after IoU matching of objects). Source: arxiv.org. Saturation forecast: Around 2028. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (xHigh) | 63 | #29 (GPT-5.6 Sol) |
| 2 | Qwen 3.5 397B A17B | 60 | #242 |
| 3 | Qwen 3.5 397B A17B (Non-reasoning) | 59 | #242 (Qwen 3.5 397B A17B) |
| 4 | GLM-5.2 | 56 | #137 |
| 5 | Kimi K2.6 (Non-reasoning) | 55 | #173 (Kimi K2.6) |
| 6 | Qwen 3.5 9B (Non-reasoning) | 52 | #803 (Qwen 3.5 9B) |
| 7 | Qwen 3.5 35B A3B (Non-reasoning) | 52 | #480 (Qwen 3.5 35B A3B) |
| 8 | MiniMax-M2.5 | 44 | #590 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=longdocbench-contextual-relationship-recovery · How It Works · Data refreshed daily, snapshot 2026-11-12.