LetTokenBurnVisual Evidence Under Cross-Examination: Evaluating and Controlling Decision-Level Evidence Use in Vision-Language Models
A benchmark of 28,000 decision problems evaluating candidate-bound visual contribution through matched invalidation and rebinding tests on vision-language models.
$ npm i -g lettokenburn-cli$ ltb bench info cross-bench
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
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct | 68.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen2.5-VL-7B | 66.3 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| LLaVA-OneVision-2-8B | 66.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 63.0 | 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.
CROSS-Bench comprises 28,000 decision roots covering visual attributes, counts, spatial relations, and chart states. It uses matched clean, invalid, and rebinding conditions to test if evidence effects follow valid supporting relations.
Scoring measures normalized effect transfer and task accuracy relative to a baseline without auxiliary evidence. Results show RIVET improves effect transfer from 0.512 to 0.651 and boosts accuracy by 5.70 pp across four backbones.