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

#ModelScore
1Qwen 3.5 0.8B (Non-reasoning)81
2Nova 2 Lite78.3
3Grok 4 Fast (Non-reasoning)76.4
4Gemini 2.5 Flash Lite74
5Qwen 3 VL 8B Instruct71.5
6Gemini 3 Flash (Minimal)71
7Qwen 3 VL 4B Instruct70.4
8Qwen 3.5 2B (Non-reasoning)65.6
9Gemini 2.5 Flash (Non-reasoning)64.7
10GPT-5 Mini (Minimal)58.8
11GPT-5 Nano (Minimal)46.6
12Ministral-3-3B-Instruct-251245.6
13Qwen 3.5 4B (Non-reasoning)44.6
14Gemma 4 E2B43.4
15Ministral-3-8B-Instruct-251243.4

Interactive version: theaggregate.ai/benchmark?slug=prism-vlm-sycophancy-resistance · How It Works · Data refreshed daily, snapshot 2026-09-26.