LetTokenBurnOmniVCBench: Benchmarking Evidence-Grounded Multimodal Reasoning Towards AI Virtual Cells
A figure-centric benchmark with 6,077 curated question-answer pairs evaluating evidence-grounded multimodal reasoning (inference, explanation, hypothesis) in scientific figures.
$ npm i -g lettokenburn-cli$ ltb bench info omnivcbench
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 | |||
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 55.1 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Grok-4.6 | 50.3 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Claude Sonnet 4.5 (2025-09-29) | 49.4 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| GPT-5.6 Luna | 43.7 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 34.3 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Qwen3-VL-4B | 30.5 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 30.3 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| InternVL3.5-4B | 28.9 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Qwen2.5-VL-7B | 26.6 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| LLaVA-OneVision-7B | 25.5 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Qwen2.5-VL-3B | 22.5 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Qwen3-VL-2B | 21.7 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| InternVL3.5-2B | 20.9 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| SmolVLM2-2.2B | 17.3 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| LLaVA-Med-Mistral-7B | 15.9 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Content & Construction: OmniVCBench is a source-traceable benchmark for the interpretation component of Artificial Intelligence Virtual Cells. It contains 6,077 single- and multi-subfigure question-answer pairs derived from cellular research figures and experimental contexts (e.g., microscopy fields, immunoblots) in scientific literature.
Scoring & Key Findings: The evaluation uses paired open-response scoring via an AIVC-Judge framework with reference-aware rubrics and controlled MCQ accuracy. Proprietary models reached only 3.28/5.00 on the judge scores, indicating significant room for improvement in interpreting experimental evidence.