IAAILoopLook
Loop to Look Better: Learning Adaptive Depth for Question-Guided Perception in Vision-Language Models
Adds adaptive recurrent perception to pretrained vision-language models by revisiting a middle-to-late window of the vision encoder with the question injected at each loop. A learned stopper sets the depth and a token-wise Mixture-of-Looks blends features from different loops.
Inside the method
Vision encoder: [ mid-late window + question ] × r (learned stop) → Mixture-of-Looks → LLM
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- Learned stopping per input
- KV / state strategy
- Not applicable (encoder-side)
Reading note
Across four backbones and eight benchmarks, gains reach 8.6% on MMStar and 8.1% on HallusionBench, while time to first token rises only 5.1–6.5%.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{looplook2026,
title = {Loop to Look Better: Learning Adaptive Depth for Question-Guided Perception in Vision-Language Models},
author = {Wencheng Ye and Anning Hu and Hongye Fang and Tianyi Wang and Junchi Yan},
year = {2026},
eprint = {2610.30824},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.30824}
}