PatternEval - Response-Pattern Failure Rate (Non-Thinking): leaderboard

Metric: Bad-pattern trigger rate (%, lower is better; share of final responses showing at least one of chain-of-thought leakage, repetition, logical contradiction or performative reasoning, labelled by a Seed-2.0-Pro judge with image access; non-thinking inference; PatternEval: 2,415 failure-enriched image-prompt pairs in three task families (visual perception and grounding, OCR and structured images, multimodal knowledge reasoning); official default decoding with only the reasoning control changed). Source: arxiv.org. Saturation forecast: Around December 2026. 25 models tracked.

Top models

#ModelScore
1GPT-5.5 (Non-reasoning)5.04
2GPT-5.4 (Non-reasoning)5.34
3Seed 2.0 Pro (Non-reasoning)6.94
4Claude Opus 4.822.06
5Qwen 3.7 Max (Non-reasoning)26.77
6Qwen 3.7 Plus (Non-reasoning)27.82
7Qwen 3.5 122B A10B (Non-reasoning)28.45
8Qwen 3.5 27B (Non-reasoning)29.32
9Qwen 3.5 397B A17B (Non-reasoning)29.32
10Qwen 3.6 35B A3B (Non-reasoning)32.8
11Qwen 3.6 27B (Non-reasoning)32.81
12Qwen 3.5 35B A3B (Non-reasoning)33
13Claude Opus 4.633.18
14Qwen 3 VL 32B Instruct35.03
15GPT-5 Nano (Minimal)35.67

Interactive version: theaggregate.ai/benchmark?slug=patterneval-response-pattern-failure-rate-non-thinking · How It Works · Data refreshed daily, snapshot 2026-09-29.