ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 118
EgoExo-Next
EgoExo-Next: Benchmarking Vision-Language Models on Visual-Option Next-State and Cross-View Reasoning
A visual-option benchmark with 2,503 four-choice questions evaluating dynamic visual-state reasoning, temporal progression, and cross-view correspondence in egocentric and exocentric videos.
How to run
ltb · egoexo-next
$ npm i -g lettokenburn-cli$ ltb bench info egoexo-next
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
13 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-6 | 60.2 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| Gemini 3.1 Pro | 43.8 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| GPT-5.2-chat | 43.2 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| Qwen3-VL-235B | 40.9 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| RoboBrain | 38.6 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| Qwen3-VL-32B | 37.0 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| InternVL3.5-38B | 36.8 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 36.3 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| SpatialMLLM | 35.6 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| Claude Opus 4.6 | 33.6 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| SpatialVLM | 33.0 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 32.9 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| LLaVA-OneVision-1.5-8B | 31.9 | N/A | Independent | medium | Oct 3, 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
EgoExo-Next comprises 2,503 human-curated multiple-choice questions from six public video sources. It features four subtasks: Ego Next-State, Ego-Exo Matching, Exo Next-State, and the composed Ego-to-Exo Next-State.
Scoring uses average accuracy. Evaluations reveal a significant human-model gap, with the best VLM achieving 43.81% accuracy compared to 98.55% for humans, highlighting limitations in dynamic visual-state reasoning.