Research map / Vision-language and multimodal / Multimodal LLMs

Blink: Dynamic Visual Token Resolution for Enhanced Multimodal Understanding

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Blink dynamically expands salient visual tokens via a trained convolutional TokenSR amplifier and drops them in later layers, enhancing MLLM visual perception within a single forward pass.

Problem

MLLMs have limited visual perception; existing salient-region enhancement methods rely on post-hoc selection or multiple forward passes, typically focusing on only one region.

Approach

Layer-wise attention from the last text token identifies salient patches; salient visual tokens are expanded through a learned TokenSR module (trained to match cropped-image hidden states) at layers where attention is sharp, and the expanded tokens are dropped when attention shifts, all within one forward pass.

Key results

• On LLaVA-1.5-7B, Blink improves MME Perception total by 14.02 and Cognition total by 3.93 over vanilla (MME total 1881.53 vs. ViCrop 1804.65 with 3 forward passes).
• Blink reduces FLOPs from 21.5T (ViCrop) to 9.73T with a single forward pass.
• Training-free Blink-interp still surpasses the vanilla LLaVA-1.5 on MME Perception (1514.08 vs. 1505.72).

Surprising findings

• Disabling token drop (w/o Drop) increases Blink's MME total by 2.50, contrary to the expectation that dropping is needed.
• Using 3x3 or 4x4 patch partitions does not improve over the default 2x2 setting.

Limitations

• TokenSR relies on convolution-based upsampling; alternative expansion modules are unexplored.
• Experiments are limited to models no larger than 7B parameters.
• Layer ranges and thresholds are sensitive; small shifts in layer range change MME total by up to about 7-8 points.

Experimental setup

Datasets: LLaVA-1.5 training data COCO GQA OCR-VQA TextVQA Visual Genome

Benchmarks: MME GQA MMBench MMBench-CN POPE ScienceQA MM-Vet

Metrics: accuracy MME perception score MME cognition score total score

Baselines: ViCrop Vanilla LLaVA-1.5

Builds on: ViCrop LLaVA-1.5 Visual Perception Token

Closest papers by meaning

• BLINK-Twice: You see, but do you observe? A Reasoning Benchmark on Visual Perception (NeurIPS 2025 · similarity 0.90)
• See What You Are Told: Visual Attention Sink in Large Multimodal Models (ICLR 2025 · similarity 0.90)
• Perception Tokens Enhance Visual Reasoning in Multimodal Language Models (CVPR 2025 · similarity 0.89)
• Seeing What Matters: A Training-Free Self-Guided Framework for Multimodal Detail Perception and Reasoning (CVPR 2026 · similarity 0.89)
• Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · similarity 0.89)
• What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models (CVPR 2026 · similarity 0.89)