NightIQA-Bench - Reasoning: leaderboard
Nighttime image-quality understanding MCQs across perception, technical assessment, localization and reasoning. MCQ accuracy (%). NightIQA-Bench v1 Table2,1078held-out MCQs from816scene-disjoint images, all0–100accuracy. Zero-shot baseline protocol; local SFT+GRPO8B is a separately trained named system. Native source does not resolve every instruct/thinking checkpoint or closed-model snapshot; those remain source scoped. Sampled random guess is a reference, not an evaluated model.
Metric: MCQ accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B (Benchmark Sweep Run: source=NightIQA; zero-shot; native Table2 system; checkpoint/effort unreported) | 82.47 | |
| 2 | Qwen 3.5 35B A3B (Benchmark Sweep Run: source=NightIQA; zero-shot; native Table2 system; checkpoint/effort unreported) | 80.76 | |
| 3 | NightIQA SFT+GRPO 8B (Benchmark Sweep Run: source=NightIQA; zero-shot; native Table2 system; checkpoint/effort unreported) | 80.07 | |
| 4 | Gemini 3.1 Flash Lite (Benchmark Sweep Run: source=NightIQA; zero-shot; native Table2 system; checkpoint/effort unreported) | 79.38 | |
| 5 | Qwen 3 VL 235B A22B (Benchmark Sweep Run: source=NightIQA; zero-shot; native Table2 system; checkpoint/effort unreported) | 77.66 | |
| 6 | GPT-4o (Benchmark Sweep Run: source=NightIQA; zero-shot; native Table2 system; checkpoint/effort unreported) | 75.95 | |
| 7 | GPT-5 Mini (Benchmark Sweep Run: source=NightIQA; zero-shot; native Table2 system; checkpoint/effort unreported) | 75.26 | |
| 8 | Qwen 3.5 122B A10B (Benchmark Sweep Run: source=NightIQA; zero-shot; native Table2 system; checkpoint/effort unreported) | 75.26 | |
| 9 | Qwen 3.5 27B (Benchmark Sweep Run: source=NightIQA; zero-shot; native Table2 system; checkpoint/effort unreported) | 74.91 | |
| 10 | MiMo-V2.5 (Benchmark Sweep Run: source=NightIQA; zero-shot; native Table2 system; checkpoint/effort unreported) | 74.57 |
Interactive version: theaggregate.ai/benchmark?slug=nightiqa-bench-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-06.