ChartAnno - Intent-Level Semantic Consistency: leaderboard
ChartAnno (2026) asks multimodal LLMs to add annotations to the executable code of real-world charts from instructions at three levels of specificity. A GPT-5.4 judge rates semantic faithfulness and clarity from 1 to 5 (0 for failed output); higher is better. Chart code is the only input, and intent-level instructions state only the communicative goal.
Metric: Semantic consistency (out of 5): mean of semantic faithfulness and semantic clarity, each an integer rubric score from 1 to 5 given by a GPT-5.4 judge, 0 for failed or missing output; annotated chart code generated from the unannotated code of 1,200 real-world charts; chart code input only; intent-level instructions that state only the communicative intent. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 3.69 | #91 |
| 2 | Gemini 3 Flash (Preview) | 3.59 | #141 |
| 3 | GPT-5.4 | 3.56 | #121 |
| 4 | Claude Sonnet 4.6 | 3.49 | #158 |
| 5 | Kimi K2.5 | 3.32 | #252 |
| 6 | Gemma 4 31B | 3.2 | #393 |
| 7 | Qwen 3.5 397B A17B | 3.08 | #242 |
| 8 | Qwen 3.5 27B | 2.97 | #358 |
| 9 | Qwen 3.5 122B A10B | 2.93 | #329 |
| 10 | Qwen 3.5 9B | 2.31 | #803 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=chartanno-intent-level-semantic-consistency · How It Works · Data refreshed daily, snapshot 2026-11-12.