OI-Bench - Instructional Interference - Injected Accuracy: leaderboard
OI-Bench (Liou et al., 2026) accuracy (%) on its 3,000 zero-shot four-choice questions after a fifth, question-irrelevant option is appended whose instructional interference directive (flattery, self-contradictory instructions, fake chain-of-thought tags and a direct command) urges its selection; averaged over that category's directive types. Higher is better.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 88.43 | #163 |
| 2 | Claude Haiku 4.5 | 85.68 | #568 |
| 3 | Llama 4 Maverick | 79.12 | #1105 |
| 4 | Gemini 2.5 Pro | 77.9 | #262 |
| 5 | Grok 4.1 | 76.71 | #473 |
| 6 | Llama 4 Scout | 76.05 | #1787 |
| 7 | Qwen 3 235B A22B | 73.99 | #622 |
| 8 | GPT-5 Mini | 73.75 | #349 |
| 9 | DeepSeek R1 | 70.12 | #467 |
| 10 | DeepSeek V3.2 | 68.7 | #403 |
| 11 | Gemini 2.5 Flash Lite | 64.05 | #958 |
| 12 | Qwen 3 8B | 63.93 | #1838 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=oi-bench-instructional-interference-injected-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.