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
VisionQ
VisionQ: VLM-as-a-Judge Taxonomy, Dataset and Benchmark for Qualitative Analysis in Computer Vision
A benchmark of 3,911 criterion-conditioned visual discrimination questions from 1,800+ CV comparison figures to evaluate VLM judges on specific visual criteria.
How to run
ltb · visionq
$ npm i -g lettokenburn-cli$ ltb bench info visionq
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 | |||
|---|---|---|---|---|---|---|
| OpenAI GPT-5.3-codex | 63.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| OpenAI GPT-5.5 | 63.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| OpenAI GPT-5.2 | 61.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| OpenAI GPT-5.4 | 61.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GLM-5V Turbo | 59.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| OpenAI GPT-5.1 | 59.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Xiaomi MiMo v2.5 | 59.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Anthropic Opus 4.6 | 58.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Qwen3.6 35B A3B | 58.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| google/gemma-4-31B-it | 57.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Grok-4.3 | 56.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Qwen3.6 27B | 55.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Anthropic Sonnet 4.6 | 54.0 | N/A | Independent | medium | Sep 30, 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
VisionQ comprises 3,911 hand-annotated data points linking method crops to author-stated visual claims from 1,409 peer-reviewed CV papers. It uses a six-axis taxonomy of visual criteria to ground judgments.
The evaluation protocol hides method names and reports accuracy per criterion. Strongest VLM judges reach only 63.1% accuracy, revealing significant reliability variations across different visual criteria.