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
MVGraph
MVGraph: Benchmarking LVLMs on Evidence Composition across Multi-View Visual Graph
A benchmark with 72k instances evaluating LVLMs on multi-graph reasoning across partial and temporal visual views, decomposing performance into perception, alignment, composition, and reasoning capabilities.
How to run
ltb · mvgraph
$ npm i -g lettokenburn-cli$ ltb bench info mvgraph
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
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GLM-5.3 Flash | 52.4 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Qwen3.8 Flash | 50.4 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Gemma-4-12B-it | 20.5 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| DeepSeek Flash | 17.4 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| qwen3.5-9B | 13.5 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Ovis2.5-9B | 7.8 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| GLM-4.6V-Flash | 7.4 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Molmo2-8B | 5.7 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 5.4 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Phi-4-reasoning-vision-15B | 5.4 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 5.4 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| LLaVA-OneVision-2-8B | 5.3 | N/A | Independent | medium | Oct 9, 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
MVGraph comprises 72,000 programmatically constructed instances spanning two multi-graph scenarios (partial and temporal) and 37 tasks. It ensures no single view suffices for answers, requiring cross-view evidence composition.
The benchmark uses a capability-oriented diagnostic schema with golden intermediate representations to measure Perception, Alignment, Composition, and Reasoning. Evaluations reveal cross-view edge alignment and relation-level composition as major bottlenecks for current LVLMs.