CDR-Bench - Order-Sensitive Filter Consistency: leaderboard
Group-level order consistency on CDR-Bench (2026), a benchmark of natural-language data-refinement recipes: the percent of filter-order-sensitive recipe groups for which a model gets every permutation of the same operators right (filter placed before, between and after the mappers) within three attempts, in non-thinking mode. Higher is better.
Metric: Group-level order consistency OCS@3 (%) on the filter-order-sensitive recipe groups of CDR-Bench: a group counts only when the model gets every permutation of the same operators right (filter placed before, between and after the mappers) within 3 attempts; non-thinking mode, baseline prompt; higher is better. Source: arxiv.org. Saturation forecast: Around December 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 18.68 | #102 |
| 2 | GPT-5.4 (Non-reasoning) | 9.46 | #123 (GPT-5.4) |
| 3 | Qwen 3.6 Max | 8.27 | #180 |
| 4 | Claude Opus 4.5 | 8.14 | #156 |
| 5 | Qwen 3.6 27B (Non-reasoning) | 7.09 | #341 (Qwen 3.6 27B) |
| 6 | GLM-5 (Non-reasoning) | 6.47 | #260 (GLM-5) |
| 7 | Gemma 4 31B (IT) (Non-reasoning) | 6.11 | #342 (Gemma 4 31B (IT)) |
| 8 | Kimi K2.6 (Non-reasoning) | 5.88 | #178 (Kimi K2.6) |
| 9 | Qwen 3.6 35B A3B (Non-reasoning) | 3.7 | #462 (Qwen 3.6 35B A3B) |
| 10 | DeepSeek V4 Flash (Non-reasoning) | 2.4 | #258 (DeepSeek V4 Flash) |
| 11 | Llama 4 Scout | 0.86 | #1776 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cdr-bench-order-sensitive-filter-consistency · How It Works · Data refreshed daily, snapshot 2026-11-13.