Paper 2607.24957
PerceptionBench: Evaluating Atomic Visual Perception in Multimodal Large Language Models
- Published
- Jul 2026
- Research lab
- Moonshot AI
- Citations
- 0
- GitHub
- 170 stars
01 In brief
Summary
PerceptionBench is a benchmark introduced by Moonshot AI to evaluate atomic visual perception in Multimodal Large Language Models (MLLMs).
It addresses limitations of existing benchmarks that conflate perception with reasoning or knowledge.
The benchmark was constructed bottom-up: failures of frontier MLLMs on 42 existing benchmarks were attributed to their earliest erroneous step, yielding an error taxonomy whose perception branch defines ten atomic capabilities (e.g., localization, counting, OCR, hallucination).
PerceptionBench contains 3,000 verified questions, each isolating a single capability, with difficulty stemming from perception.
Evaluations of sixteen frontier MLLMs show no model exceeds 60% accuracy; perception-related hallucination is the weakest capability on average.
Models with similar overall scores show divergent capability profiles, and per-sample behavior is unstable across runs.
The open-source model Kimi K3 trails the leader by only 1.2 points.
02 From the paper
Abstract
We introduce PerceptionBench, a benchmark specifically designed to evaluate the atomic visual perception capabilities of Multimodal Large Language Models (MLLMs). Existing benchmarks often fail to isolate perception: holistic evaluations conflate perceptual errors with failures in reasoning or domain knowledge, while application-driven benchmarks only cover narrow, fragmented domains shaped by heuristic designs. To address these limitations, PerceptionBench adopts a bottom-up approach: by diagnosing the earliest failure points in the responses of frontier MLLMs across 42 existing benchmarks, we construct an error taxonomy whose perception branch defines ten atomic perceptual capabilities. Guided by this taxonomy, we construct 3,000 verified questions with short, unambiguous answers, each isolating a single capability, with difficulty stemming from perception rather than reasoning or knowledge. Benchmark results across sixteen frontier MLLMs reveal that atomic perception remains largely unsolved---no model reaches 60\% accuracy, perception-related hallucination is the weakest capability on average, and similar overall scores conceal sharply divergent capability profiles. PerceptionBench thus provides a capability-level standard for measuring and diagnosing the visual perception boundaries of MLLMs.