ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 118
EMBER-Bench
EMBER-Bench: Benchmarking Cross-Event Causal Memory in Long-Horizon Embodied Tasks
An egocentric benchmark with 189 household tasks and 699 QA pairs evaluating cross-event causal memory via next-action prediction and causal traceback in long-horizon videos.
How to run
ltb · ember-bench
$ npm i -g lettokenburn-cli$ ltb bench info ember-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
Scores
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Gemini 3.8 Flash | 72.8 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Qwen 3.8 Omni Flash | 55.6 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Qwen3.5 397B A17B | 53.6 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Qwen3.8 27B | 43.0 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Qwen3.8 Flash | 42.4 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Qwen3.5 122B A10B | 37.1 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Gemini 3.1 Pro | 26.8 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Qwen3.8 Max | 26.6 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Qwen3-VL-235B-A22B | 26.5 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| Qwen3-VL-32B | 26.5 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| MiMo V2.6 Pro | 26.3 | N/A | Independent | medium | Oct 4, 2026 | arxiv.org |
| GPT-6 Sol | 25.7 | N/A | Independent | medium | Oct 4, 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
Contains 189 household tasks and 699 QA pairs across four memory levels (progress, anomalies, interventions, compound). It uses pre-recorded egocentric videos with fine-grained event annotations.
Evaluates next-action prediction and causal traceback. Best model accuracy is 61.2%, significantly below human performance (98.3%), highlighting difficulties in causal reasoning.