Efficient vision language models improved with self distillation and token pruning
SCOPD: Sparse-Context On-Policy Self-Distillation for Efficient Vision-Language Models
Computer Vision and Pattern Recognition
Summary
Vision-language models that understand images and videos use many pieces of visual information, which makes them slow and costly to run. Cutting down the number of visual pieces speeds things up but can cause the model to miss important details. The authors found that sometimes the model still has the needed information but doesn't use it well. They introduced a training method called SCOPD that helps the model learn to use fewer visual pieces more reliably without extra cost during use. This approach keeps more accuracy compared to usual pruning methods, even with very limited visual data.
What this means in practice
- •For mobile app developers: Create apps that understand images and videos faster by integrating efficient vision-language models trained with SCOPD to use fewer visual data points.
- •For cloud service engineers: Reduce computational costs for large-scale visual understanding services by adopting SCOPD to maintain accuracy with pruned visual inputs.
Authors
Ahmadreza Jeddi, Enming Zhang, Jasper Gerigk, Hakki Karaimer, Mozhgan Nasr Azadani, Jiayun Luo, Minh Ngoc Le, Gholamali Aminian, Hugo Buurmeijer, Yongchao Chen, Leonid Sigal, Igor Gilitschenski, Konstantinos G. Derpanis, Marco Pavone, Babak Taati
Abstract
Reasoning vision-language models (VLMs) process images and videos as long sequences of visual tokens, making inference expensive. Training-free token pruning reduces this cost, but aggressive compression can sharply degrade performance, often attributed to irreversible loss of task-relevant visual information. We show that this explanation is incomplete. In a fixed-context Pass@K analysis, repeated sampling from the same pruned visual representation recovers many examples missed by greedy decoding, indicating that useful visual evidence can remain accessible but be used unreliably. We call this the representation-utilization gap. Motivated by this observation, we introduce SCOPD, a sparse-context on-policy self-distillation framework in which a student generates reasoning trajectories from pruned visual tokens while a privileged full-context teacher supervises the same on-policy prefixes. SCOPD requires no ground-truth responses, architectural changes, or additional inference-time computation. We further introduce SCOPD+, which uses a small visual-budget intervention to identify visually sensitive response positions and selectively distill them. At 10% visual-token retention, the Vanilla model retains 86.37% of its unpruned performance across 13 benchmarks. SCOPD raises this to 90.49%, while SCOPD+ further improves it to 92.43%. Across token budgets, benchmarks, and pruning operators, our results show that efficient reasoning depends not only on which visual information survives pruning, but also on how reliably the model learns to use it.