HM-Bench (VSR2): leaderboard

HM-Bench (2026) tests multimodal LLMs on hyperspectral image understanding with 19,337 multiple-choice questions over 2,178 samples in 13 perception and reasoning tasks, given full VSR2 input: the RGB, PCA and report views together with the authors' fixed cross-view fusion prompt. Accuracy (%), zero-shot; higher is better.

Metric: Average accuracy (%) over all tasks of HM-Bench's 19,337 hyperspectral multiple-choice questions (2,178 samples, 13 task categories) with full VSR2 input: the RGB, PCA and report views together with the authors' fixed cross-view fusion prompt; zero-shot, one option letter as output, temperature 0, at most 64 new tokens; higher is better. Source: arxiv.org. Saturation forecast: Around 2032. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1InternVL3.5-14B45.09#1519
2GPT-5.4 Mini44.28#409
3Claude Sonnet 4.644.01#179
4GLM-4.6V-Flash43.61#1049
5Qwen 3 VL 4B43.23#1163
6GeoLLaVA-8K41.94
7LLaVA-v1.5-7B40.6#7642
8InternVL3-9B40.47#1746
9Kimi-VL-A3B-Instruct40.36#2272
10InternVL3-14B40.05#1227
11Llama-3.1-Nemotron-Nano-VL-8B-V139.61
12GeoChat39.6#3175
13Grok 439.5#321
14Qwen 2.5 VL 7B Instruct39.41#1919
15LLaVA-Next-8B39.28#2058

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=hm-bench-vsr2 · How It Works · Data refreshed daily, snapshot 2026-11-22.