AgentHorizon
Harbor: pendingA benchmark of 1,373 computer-use tasks with paired instructions to evaluate agentic judges on long-horizon trajectory success verification across three OSs.
- Samples
- 1,373
- Released
- 2026
- Scores
- 10
FigCodeBench
From Pixel to Coding: Evaluating the Figure Reproduction Capabilities of MLLMs
A benchmark with 6,194 instances evaluating MLLM figure reproduction across 7 categories and 4 languages (Python, Matlab, R, Latex) using visual and code metrics.
How to run
ltb · figcodebench
$ npm i -g lettokenburn-cli$ ltb bench info figcodebench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
13 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Gemini 3.1 Pro | 83.7 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| GPT-5.4 | 79.7 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Gemini 3 Flash-TK | 78.5 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Gemini 3 Flash Preview | 76.8 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| GPT-5.2-chat | 76.3 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Claude Opus 4.7 | 76.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| GPT-5.4-mini | 75.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen3.6 Plus | 73.9 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Grok-4.3 | 69.5 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Gemini 2.5 Flash-TK | 58.8 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Gemini 2.5 Flash | 55.8 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| GLM-4.6V | 50.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| GLM-5.1 | 44.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
FigCodeBench contains 6,194 figure-code pairs across Python, Matlab, R, and Latex, covering complex scientific figures. It uses a multi-dimensional protocol assessing visual fidelity (PSNR, SSIM) and syntactic isomorphism.
The benchmark reveals a universal performance cliff across models and languages, with Latex posing a significant bottleneck. It highlights gaps in 2D spatial reasoning and abstract set relationship representation in current MLLMs.