ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 120
ChemExplain
ChemExplain: A Multimodal Benchmark for Explanatory Chemical Reasoning
A multimodal benchmark of 1,180 multiple-choice questions evaluating explanatory attribution in chemistry by linking molecular images to observed effects.
How to run
ltb · chemexplain
$ npm i -g lettokenburn-cli$ ltb bench info chemexplain
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 | |||
|---|---|---|---|---|---|---|
| GPT-5.4 | 78.6 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Claude Sonnet 4.6 | 78.2 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Gemini 3.1 Pro | 78.0 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| GPT-5 | 77.4 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| GPT-4o | 68.1 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Qwen3-VL | 58.8 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Phi-4-reasoning-vision | 51.3 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| InternVL3.5 | 48.0 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Qwen2.5-VL Instruct | 42.6 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| ChemVLM | 34.0 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Janus-Pro | 23.2 | N/A | Independent | medium | Oct 12, 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
ChemExplain comprises 1,180 multiple-choice questions derived from 383 peer-reviewed papers. Each item pairs a molecular or reaction image with an observed effect and four candidate explanations tied to visible structural features.
Constructed via a multi-agent framework with expert validation, it measures explanatory attribution. The strongest model achieved 78.6% accuracy, with attribution failures accounting for the majority of errors.