PatternEval - Response-Pattern Failure Rate (Thinking): leaderboard
PatternEval (2026) collects 2,415 multimodal prompts on which model responses often fail in user-visible ways, and runs hybrid-thinking models once with thinking on and once with it off. This board is thinking inference; the score is the share of final responses showing chain-of-thought leakage, repetition, contradiction or performative reasoning, as labelled by a Seed-2.0-Pro judge, in percent; lower is better.
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; 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: Estimated already saturated. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 0.78 | #49 |
| 2 | Kimi K2.6 (Thinking) | 0.99 | #178 (Kimi K2.6) |
| 3 | GPT-5.4 (Thinking) | 2.48 | #123 (GPT-5.4) |
| 4 | Claude Opus 4.8 (Thinking) | 3.55 | #56 (Claude Opus 4.8) |
| 5 | Claude Opus 4.6 (Thinking) | 3.59 | #102 (Claude Opus 4.6) |
| 6 | Qwen 3.5 397B A17B (Thinking) | 4.23 | #246 (Qwen 3.5 397B A17B) |
| 7 | Qwen 3 VL 32B (Thinking) | 4.82 | #559 |
| 8 | Qwen 3.5 27B (Thinking) | 5.31 | #359 (Qwen 3.5 27B) |
| 9 | Qwen 3.7 Max (Thinking) | 5.56 | #145 (Qwen 3.7 Max) |
| 10 | Qwen 3.5 122B A10B (Thinking) | 5.86 | #331 (Qwen 3.5 122B A10B) |
| 11 | Qwen 3.6 35B A3B (Thinking) | 5.9 | #462 (Qwen 3.6 35B A3B) |
| 12 | Qwen 3.6 27B (Thinking) | 5.93 | #341 (Qwen 3.6 27B) |
| 13 | Kimi K2.5 (Thinking) | 6.14 | #256 (Kimi K2.5) |
| 14 | Qwen 3.5 35B A3B (Thinking) | 6.37 | #482 (Qwen 3.5 35B A3B) |
| 15 | Qwen 3 VL 30B A3B (Thinking) | 10.66 | #712 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=patterneval-response-pattern-failure-rate-thinking · How It Works · Data refreshed daily, snapshot 2026-11-13.