Persistent Visual Memory: Sustaining Perception for Deep Generation in LVLMs

TL;DR

This paper introduces Persistent Visual Memory (PVM), a lightweight module that enhances sustained visual perception in LVLMs, improving accuracy and robustness in complex reasoning tasks with minimal overhead.

Contribution

The paper proposes PVM, a novel parallel module for LVLMs that mitigates visual signal decay and improves visual perception during deep generation.

Findings

PVM improves accuracy across 4B and 8B LVLMs.

PVM enhances robustness in longer generations.

PVM accelerates internal prediction convergence.

Abstract

While autoregressive Large Vision-Language Models (LVLMs) demonstrate remarkable proficiency in multimodal tasks, they face a "Visual Signal Dilution" phenomenon, where the accumulation of textual history expands the attention partition function, causing visual attention to decay inversely with generated sequence length. To counteract this, we propose Persistent Visual Memory (PVM), a lightweight learnable module designed to strengthen sustained, on-demand access to visual evidence. Integrated as a parallel branch alongside the Feed-Forward Network (FFN) in LVLMs, PVM establishes a distance-agnostic retrieval pathway that directly provides visual embeddings for enhanced visual perception, thereby structurally mitigating the signal suppression inherent to deep generation. Extensive experiments on Qwen3-VL models demonstrate that PVM brings notable improvements with negligible parameter overhead, delivering consistent average accuracy gains across both 4B and 8B scales, particularly in complex reasoning tasks that demand persistent visual perception. Furthermore, in-depth analysis reveals that PVM shows improved robustness in longer generations and accelerates internal prediction convergence.