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
VisualWorldBench
VisualWorldBench: A Fine-Grained Multi-Task Benchmark for Evaluating Visual World Knowledge in MLLMs
A taxonomy-grounded benchmark with 32,541 instances across 8 domains and 4 tasks (recognition, selection, localization, naming) to evaluate fine-grained visual world knowledge in MLLMs.
How to run
ltb · visualworldbench
$ npm i -g lettokenburn-cli$ ltb bench info visualworldbench
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 | 80.4 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Doubao Seed 2.0 Pro | 73.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Kimi K2.5 | 62.9 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Mimo-V2-Omni | 56.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| qwen3.5-27B | 53.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.4 | 51.2 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.2-chat | 47.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| google/gemma-4-31B-it | 47.4 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| qwen3.5-9B | 46.6 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 40.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| google/gemma-4-26B-A4B-it | 36.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3.5 4B | 36.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3-VL-4B | 33.7 | N/A | Independent | medium | Oct 14, 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
VisualWorldBench comprises 32,541 evaluation instances covering 15,479 categories across 8 visual domains, built on a hierarchical taxonomy of 21,291 leaf-node classes. It employs four complementary tasks: closed-set recognition, contrastive selection, precise localization, and open-world naming.
Scoring relies on accuracy across these tasks. Experiments show the strongest closed-source model achieves only 80.4% average accuracy, revealing significant capability gaps in fine-grained entity discrimination compared to conventional benchmarks.