CodeShrink: Adaptive Visual Compression for Efficient Multimodal Code Understanding

Researchers introduced CodeShrink, an adaptive visual compression framework for code understanding.

It combines blank-free rendering, reinforcement-learning-based configuration selection, and instruction-aware token pruning.

Across code QA, clone detection, and code completion, it cut visual token usage by up to 71.2% while matching or exceeding baseline performance.

The method could make multimodal large language models for code more efficient and practical without sacrificing accuracy.