LetTokenBurnWorldFact-Bench: Beyond Image-Internal Plausibility to Image-World Consistency
Evaluates image-world consistency using 1,274 source-aligned real-fake pairs across four verification regimes and ten semantic domains.
$ npm i -g lettokenburn-cli$ ltb bench info worldfact-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 | |||
|---|---|---|---|---|---|---|
| Gemini-3.5-flash | 32.2 | direct judgment | Independent | medium | Oct 8, 2026 | arxiv.org |
| MiniMax-M3 | 16.8 | direct judgment | Independent | medium | Oct 8, 2026 | arxiv.org |
| PROBE | 14.2 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Kimi K2.6 | 12.9 | direct judgment | Independent | medium | Oct 8, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 8.9 | direct judgment | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 8.1 | direct judgment | Independent | medium | Oct 8, 2026 | arxiv.org |
| Effort | 7.1 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GLM-4.6V | 3.6 | direct judgment | Independent | medium | Oct 8, 2026 | arxiv.org |
| AIDE | 2.6 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| CO-SPY | 1.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| VERITAS | 0.6 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| FakeVLM | 0.0 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| UniGenDet | 0.0 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
The benchmark contains 1,274 source-aligned real–fake image pairs designed to test factual conflicts while preserving visual plausibility. It covers four verification regimes and ten semantic domains.
Scoring uses pair accuracy, requiring both images in a pair to be classified correctly. Experiments show strong label biases in detectors, with the best direct judgment baseline achieving only 32.20% pair accuracy.