TraversalBench - Reading Order: leaderboard
TraversalBench (2026) auxiliary reading-order benchmark: vision-language models read out the vertex labels of a path laid out in one of four scan orders, testing whether layout priors such as left-to-right reading bias traversal. Token-level accuracy in percent, macro-averaged over layout conditions, higher is better.
Metric: Token-level accuracy (%): position-wise share of correctly ordered vertex labels after standardized parsing, when a model recovers the vertex sequence of a polyline drawn in one of four scan-order regimes (left-to-right or right-to-left then top-to-bottom, top-to-bottom then left-to-right or right-to-left), macro-averaged over lane-count and vertex-count combinations within each regime; shared prompt template and output format; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 98.7 | #135 |
| 2 | Gemini 3.1 Flash Lite | 90.1 | #371 |
| 3 | Qwen 3 VL 32B | 67.8 | #622 |
| 4 | Claude Opus 4.6 | 63.9 | #113 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=traversalbench-reading-order · How It Works · Data refreshed daily, snapshot 2026-11-21.