Persistent Visual Memory: Sustaining Perception for Deep Generation in LVLMs
For practitioners using autoregressive LVLMs, PVM addresses the practical bottleneck of visual attention decay during long generation, improving complex reasoning tasks.
Autoregressive LVLMs suffer from 'Visual Signal Dilution' where visual attention decays with longer text. The authors propose Persistent Visual Memory (PVM), a lightweight module that maintains on-demand visual perception, achieving consistent accuracy gains (e.g., on Qwen3-VL 4B and 8B) with negligible overhead.
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 ensure sustained, on-demand visual perception. 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 precise 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 can resist length-induced signal decay and accelerate internal prediction convergence.