PRISM-VLM - Sycophancy Resistance: leaderboard
The sycophancy-resistance axis of PRISM-VLM (NAVER, 2026): after a correct first answer the model faces three turns of pressure, a request to reconsider, an expression of doubt and a plausible wrong hint. Share of initially correct answers kept through all three turns, in percent; higher is better.
Metric: Per-axis mean score (%; share of turn-0-correct answers held through three turns of pressure ending in a plausible wrong hint; over PRISM-VLM's 6,238 items recycled from 15 public VLM benchmarks (five seeds of 100 items per benchmark); GPT-5 (low effort) synthesizes the perturbations and grades the open-ended axes; each model at the lowest reasoning effort its provider exposes, temperature 0). Source: arxiv.org. Saturation forecast: Not forecast. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 0.8B (Non-reasoning) | 81 | #8139 (Qwen 3.5 0.8B) |
| 2 | Nova 2.0 Lite | 78.3 | #1133 |
| 3 | Grok 4 Fast (Non-reasoning) | 76.4 | #438 (Grok 4 Fast) |
| 4 | Gemini 2.5 Flash Lite | 74 | #958 |
| 5 | Qwen 3 VL 8B Instruct | 71.5 | #982 |
| 6 | Gemini 3 Flash (Minimal) | 71 | #159 (Gemini 3 Flash) |
| 7 | Qwen 3 VL 4B Instruct | 70.4 | #1281 |
| 8 | Qwen 3.5 2B (Non-reasoning) | 65.6 | #4849 (Qwen 3.5 2B) |
| 9 | Gemini 2.5 Flash (Non-reasoning) | 64.7 | #452 (Gemini 2.5 Flash) |
| 10 | GPT-5 Mini (Minimal) | 58.8 | #349 (GPT-5 Mini) |
| 11 | Gemini 3.1 Flash Lite (Minimal) | 54.1 | #363 (Gemini 3.1 Flash Lite) |
| 12 | GPT-5 Nano (Minimal) | 46.6 | #888 (GPT-5 Nano) |
| 13 | Ministral-3-3B-Instruct-2512 | 45.6 | #2592 |
| 14 | Qwen 3.5 4B (Non-reasoning) | 44.6 | #1180 (Qwen 3.5 4B) |
| 15 | Ministral-3-8B-Instruct-2512 | 43.4 | #1866 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=prism-vlm-sycophancy-resistance · How It Works · Data refreshed daily, snapshot 2026-11-12.