CROSS-Bench
Harbor: pendingA benchmark of 28,000 decision problems evaluating candidate-bound visual contribution through matched invalidation and rebinding tests on vision-language models.
- Samples
- 28,000
- Released
- 2026
- Scores
- 4
RoadSceneVQA-G
Answer with Evidence: Consistency-Aware Grounded Visual Question Answering for Roadside Traffic Scenes
A benchmark of 34.7K question-answer pairs for roadside traffic scenes where every free-form answer is linked to witnessing bounding boxes to measure answer-grounding consistency.
How to run
ltb · roadscenevqa-g
$ npm i -g lettokenburn-cli$ ltb bench info roadscenevqa-g
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
11 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GroundingDINO-T | 72.7 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Groma | 61.8 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 52.2 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| qwen3.5-27B | 49.1 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| DINO-X | 40.2 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 36.9 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| YOLO-World-v2-L | 19.8 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Shikra | 18.3 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Ferret | 15.7 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| MiniCPM-V 4.6 | 13.8 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| LLaVA-OneVision-2-8B | 8.7 | N/A | Independent | medium | Oct 4, 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
RoadSceneVQA-G contains 34.7K open-ended questions on roadside traffic imagery, with each answer paired with zero or more ground-truth bounding boxes that witness the claim. It addresses say-point mismatch by requiring explicit alignment between textual assertions and visual evidence.
The benchmark uses the Answer-Grounding Consistency (AGC) evaluation suite to penalize contradictions between answers and localized boxes. Experiments show that existing models suffer from low consistency, which the proposed EtA method significantly improves.