PRISM-VLM - Sycophancy Resistance: leaderboard
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: Around May 2028. 42 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.5 0.8B (Non-reasoning) | 81 |
| 2 | Nova 2 Lite | 78.3 |
| 3 | Grok 4 Fast (Non-reasoning) | 76.4 |
| 4 | Gemini 2.5 Flash Lite | 74 |
| 5 | Qwen 3 VL 8B Instruct | 71.5 |
| 6 | Gemini 3 Flash (Minimal) | 71 |
| 7 | Qwen 3 VL 4B Instruct | 70.4 |
| 8 | Qwen 3.5 2B (Non-reasoning) | 65.6 |
| 9 | Gemini 2.5 Flash (Non-reasoning) | 64.7 |
| 10 | GPT-5 Mini (Minimal) | 58.8 |
| 11 | GPT-5 Nano (Minimal) | 46.6 |
| 12 | Ministral-3-3B-Instruct-2512 | 45.6 |
| 13 | Qwen 3.5 4B (Non-reasoning) | 44.6 |
| 14 | Gemma 4 E2B | 43.4 |
| 15 | Ministral-3-8B-Instruct-2512 | 43.4 |
Interactive version: theaggregate.ai/benchmark?slug=prism-vlm-sycophancy-resistance · How It Works · Data refreshed daily, snapshot 2026-09-26.