# BLINK ![Category](https://img.shields.io/badge/Category-Reasoning-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Accuracy-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-Apr%202024-555) [![Paper](https://img.shields.io/badge/Paper-arXiv-b31b1b)](https://arxiv.org/abs/2404.12390) [![GitHub](https://img.shields.io/badge/GitHub-zeyofu%2FBLINK__Benchmark-181717?logo=github&logoColor=white)](https://github.com/zeyofu/BLINK_Benchmark) [![Dataset](https://img.shields.io/badge/Dataset-HuggingFace-ff9d00?logo=huggingface&logoColor=white)](https://huggingface.co/datasets/BLINK-Benchmark/BLINK) > **Quick answer:** BLINK evaluates core visual perception abilities that humans can typically solve "within a blink" — such as relative depth estimation, visual similarity, spatial reasoning, and multi-view/multi-image correspondence — but which remain challenging for multimodal LLMs. LFM2.5-VL-3B scores 61.5% as of August 2026. ## At a Glance **What it tests:** 14 classic computer-vision task formats (depth estimation, relative reflectance, forensic detection, jigsaw, multi-view reasoning, visual correspondence, and more) reformatted as multiple-choice visual questions. **Why it matters:** Many perception tasks that are trivial for humans and classical computer-vision pipelines remain surprisingly hard for multimodal LLMs; BLINK isolates this gap rather than testing world knowledge or language reasoning. **Known limitations:** Because tasks are reformatted from classical CV benchmarks into multiple-choice format, chance-level performance is non-trivial (often 25-50%), so raw scores must be interpreted relative to a random baseline. ## What BLINK Measures BLINK (Fu et al., 2024) packages 14 classic visual perception tasks — drawn from established computer vision benchmarks — into a unified multiple-choice visual question-answering format. Tasks span depth estimation, relative reflectance, forensic/manipulation detection, jigsaw puzzle solving, multi-view reasoning across several images of the same scene, and visual correspondence, among others. The benchmark's premise is that these tasks are solvable by humans "within a blink" of an eye but require perception rather than reasoning or world knowledge — probing a different capability than benchmarks like MMMU or MathVista, which lean on knowledge and multi-step reasoning. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Reasoning (visual perception) | | Metric | % accuracy (multiple-choice) | | Saturation | Low | | Created by | Fu et al. | | Source paper | [Fu et al. 2024](https://arxiv.org/abs/2404.12390) | | GitHub | [zeyofu/BLINK_Benchmark](https://github.com/zeyofu/BLINK_Benchmark) | | Dataset | [BLINK-Benchmark/BLINK on HuggingFace](https://huggingface.co/datasets/BLINK-Benchmark/BLINK) | ## How BLINK Is Scored Each task is posed as a multiple-choice question over one or more images; the model's score is the percentage of questions answered correctly across the 14 task types, typically reported as an overall average. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [LFM2.5-VL-3B](/models/liquid-ai/lfm2-5-vl-3b) | 61.5% | [LFM2.5-VL-3B launch blog](https://www.liquid.ai/blog/lfm2-5-vl-3b) | 2026-08 | *Score sourced from Liquid AI's LFM2.5-VL-3B announcement, August 2026.* ## BLINK on Benchgen > No Benchgen results yet — [be the first to run BLINK](/benchmarks/fu-et-al/blink). ## BLINK vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **BLINK** | Core visual perception across 14 classic CV task formats | Low | | [MuirBench](/benchmarks/wang-et-al/muirbench) | Multi-image reasoning and comparison | Low | | [RefCOCO](/benchmarks/kazemzadeh-et-al/refcoco) | Referring expression comprehension (object grounding) | Medium | ## Run BLINK on Your Model Benchgen lets you evaluate your own vision-language model's core perception abilities and track them across model versions. ## Frequently Asked Questions

What is BLINK?

BLINK is a multimodal benchmark testing core visual perception abilities — depth, similarity, spatial reasoning, multi-view correspondence — that are easy for humans but hard for multimodal LLMs.

What does a good score look like on BLINK?

LFM2.5-VL-3B reports 61.5% as of August 2026; scores vary widely by model size and vision encoder quality.

Who created BLINK?

BLINK was created by Fu et al.; see the [original paper](https://arxiv.org/abs/2404.12390).

Is BLINK saturated?

No — BLINK remains a low-saturation benchmark, as many of its perception tasks continue to challenge even frontier multimodal models.

Rank	Model	Score
1	lfm3-vl-2b	71.1
2	lfm2-5-vl-3b	61.5
3	north-micro-vision-instruct	52.7

BLINK — Results

BLINK — Evaluations