ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 118
CogEmo-Bench
A multimodal benchmark evaluating cognitive appraisal reasoning in emotion understanding across six dimensions using the Appraisal Evidence Quality Score (AEQS).
How to run
ltb · cogemo-bench
$ npm i -g lettokenburn-cli$ ltb bench info cogemo-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
14 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen3-Omni | 52.2 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3.8 Max | 46.3 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-5.6 Luna | 43.5 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| VideoChat-Flash | 39.1 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Kimi K3 | 39.0 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-5.2-chat | 36.9 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| mPLUG-Owl3 | 34.7 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| SALMONN | 33.4 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen2-Audio | 32.7 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| AffectGPT | 31.9 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| LLaMA-VID | 31.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Chat-UniVi | 29.9 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Emotion-LLaMA | 25.9 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| PandaGPT | 18.8 | N/A | Independent | medium | Oct 8, 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
CogEmo-Bench evaluates MLLMs on perception-to-appraisal reasoning for emotion understanding. It uses a held-out subset of CogEmo-40K with over 6,000 human-verified instances covering six appraisal dimensions.
Scoring relies on the Appraisal Evidence Quality Score (AEQS) to assess cognitive-affective understanding. The benchmark reveals limitations in current models, with CogEmo-MoE achieving an AEQS of 70.60 compared to 56.30 for GPT-5.2.