VAMPS: leaderboard

Metric: Accuracy (%) on the English versions of the 218 Konkour-seed questions (Iranian university entrance exam algebra and calculus, chosen so that plotting reveals the answer; four-option multiple choice), tool-enabled visual solving: the model must obtain one to four Desmos plots and answer only from the visible graph evidence, no analytical derivation; thinking disabled for every model, temperature 0, answers read from a required JSON block; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 18 models tracked.

Top models

#ModelScore
1Claude Opus 4.793.6
2Gemma 4 31B (Non-reasoning)89.5
3Claude Sonnet 4.688.5
4Qwen 3.5 397B A17B (Non-reasoning)88.5
5Gemma 4 26B A4B (Non-reasoning)85.3
6Qwen 3.5 27B (Non-reasoning)83.5
7Gemini 2.5 Flash (Non-reasoning)78.4
8Qwen 3.5 35B A3B (Non-reasoning)76.2
9Qwen 3 VL 32B Instruct74.3
10GPT-5.4 (Non-reasoning)66.5
11Ministral 3 14B59.6
12Ministral 3 8B58.7
13Gemma 3 27B50.5
14Qwen 3 VL 8B Instruct50
15GPT-4o49.5

Interactive version: theaggregate.ai/benchmark?slug=vamps · How It Works · Data refreshed daily, snapshot 2026-09-29.