PerceptionBench finds no frontier MLLM reaches 60% on atomic visual perception
moonshotai · hf · 2026-07-29
PerceptionBench is a benchmark for atomic visual perception in multimodal LLMs. It tries to isolate perception errors from reasoning and world knowledge by building short, unambiguous questions that each test a single perceptual capability.
- Construction: derived from failure analysis across 42 existing benchmarks.
- Scale: 3,000 verified questions spanning 10 atomic perception skills.
- Main result: across 16 frontier MLLMs, no model reaches 60% accuracy.
- Finding: perception-related hallucination is the weakest capability on average, and similar total scores hide very different capability profiles.
- Value: provides a cleaner standard for diagnosing where visual perception breaks down.
Related event: Moonshot AI Releases PerceptionBench for Visual Perception(2 posts)→
More from Multimodal
- Fan animation made in Blender with Claude Code (Opus 5.5) — Icy-Appointment3583 · 2026-10-16
- Making Blender animations with Opus, Astra and Argon: how you prompt still matters — chongdashu · 2026-10-16
- Tessellation-free geometric queries on displaced surfaces cut caustics variance 10x — ssh4net · 2026-10-16
- Complex multi-link CAD mechanism finally animates, dev celebrates the breakthrough — zetalyrae · 2026-10-16
- OpenMOSS releases MOSS-Image-SFT: Apache 2.0 text-to-image model on Hugging Face — bdsqlsz · 2026-10-16
- Hyperflow open-sourced: 49 sampling steps down to 8, ~3x faster video generation — mhdfaran · 2026-10-16