GLANCE Accelerates Vision-Language Models with Lossless Speculative Decoding

Speculative decoding is a technique designed to accelerate text generation from large language models without altering the final output. However, applying this to Vision-Language Models (VLMs) has been problematic. Traditional drafters for VLMs are often small and autoregressive, meaning they cannot efficiently process image data at every step, leading to compromises like vision compression or pruning. This limitation makes the drafter less reliable precisely where visual information is most crucial for text prediction. Researchers have introduced GLANCE, the first one-pass block drafter that achieves lossless speculative decoding for unmodified VLM targets. GLANCE overcomes previous limitations by employing a block-diffusion head that directly reads the target VLM's already-fused vision-language state. This design eliminates the vision processing cost for the drafter and allows it to fill an entire block of text in a single forward pass, making depth computationally inexpensive. GLANCE verifies a wide candidate tree in one target pass, ensuring that every audited prompt exactly reproduces greedy decoding. The system particularly benefits grounded workloads, where long verbatim copies can be generated with significantly fewer passes compared to autoregressive or other block drafters. GLANCE demonstrates up to 2.93x faster decoding than autoregression and accepts 2.7x longer blocks than previous methods, all while maintaining output fidelity. The core principle governing these results is that accepted block length is determined by the target's next-token entropy, with a fitted slope that steepens with grounding across tasks.