ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 118
PhysVista
PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop
A benchmark evaluating physical intelligence in Vision-Language Models using real and AI-generated videos to test perception, reasoning (event/scale-level), and plausibility assessment.
How to run
ltb · physvista
$ npm i -g lettokenburn-cli$ ltb bench info physvista
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
4 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Doubao Seed 2.0 Pro | 91.3 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Claude Opus 5.5 | 90.7 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Kimi K2.5 | 90.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GPT-6 Sol | 88.0 | N/A | Independent | medium | Sep 30, 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
Content & Construction: PhysVista evaluates the 'seeing–reasoning–assessment' cognitive loop for VLMs. It uses both real-world and AI-generated video data to assess physical state perception, temporal/causal reasoning, and plausibility scoring.
Scoring & Findings: The framework distinguishes between event-level and scale-level reasoning tasks. Experiments reveal substantial limitations in current models regarding genuine physical understanding compared to mere visual recognition.