Kaggle Dynamic Spatial-Temporal Metacognitive Abstention: leaderboard
Community benchmark on Kaggle by rishirajacharya: text-based spatial-temporal routing problems, some of which have no valid answer. The score multiplies routing accuracy on the solvable problems by abstention accuracy on the impossible ones and scales it to 0-100, so a model must both solve and decline correctly; higher is better.
Metric: Routing x Abstention Score (0-100). Source: www.kaggle.com. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 100 | #64 |
| 2 | Claude Opus 4.6 | 98.21 | #67 |
| 3 | Gemma 4 31B | 98.21 | #212 |
| 4 | Claude Sonnet 4.6 | 94.64 | #94 |
| 5 | Gemma 4 26B A4B | 91.07 | #360 |
| 6 | Gemini 3 Flash (Preview) | 89.29 | #90 |
| 7 | GLM-5 | 87.5 | #149 |
| 8 | Gemini 2.5 Pro | 82.02 | #159 |
| 9 | GPT-OSS-20B | 80.36 | #523 |
| 10 | Gemini 3.1 Flash Lite (Preview) | 58.93 | #201 |
| 11 | Qwen 3 235B A22B 2507 Instruct | 55.36 | #315 |
| 12 | Gemini 2.5 Flash | 41.15 | #255 |
| 13 | Claude Haiku 4.5 | 39.94 | #287 |
| 14 | DeepSeek V3.2 | 33.16 | #210 |
| 15 | GPT-5.4 | 24.35 | #83 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kaggle-dynamic-spatial-temporal-metacognitive-abstention · How It Works · Data refreshed daily, snapshot 2026-11-05.