GOTS: Greedy Orthogonal Token Selection for High-Resolution Vision-Language Models

Researchers introduced GOTS, a training-free method for reducing visual tokens in high-resolution vision-language models.

GOTS greedily keeps tokens with the largest orthogonal residual energy relative to the selected set.

It shows strong results across backbones and benchmarks such as Qwen-VL, InternVL, and OCRBench, even after selection overhead is counted.

The approach lowers inference cost and latency while preserving more performance than competing token-pruning methods.