Smart glasses could help people make sense of their surroundings, but the small vision-language models that run on such devices face a tight computational budget. A new paper on arXiv proposes GazeLens, a method that uses gaze cues to selectively reduce visual tokens in egocentric video before question answering. By focusing computation where the user is actually looking, the approach aims to preserve useful evidence while dropping redundant content.
The authors identify a core difficulty: in video streams, many frames offer little new information, and processing them all consumes scarce compute. GazeLens provides a way to adaptively prune tokens based on eye movements, making the model's limited resources go further. The work is positioned as a simple solution to the efficiency problem, though details of the implementation and evaluation are not fully described in the available abstract.
As the only source, the arXiv record states the motivation and proposed direction, but does not yet include experimental results or comparisons in the released text. Whether GazeLens outperforms other token-reduction strategies, and on which benchmarks, would require reading the full paper or future coverage. For now, the contribution is a practical idea: let the wearer's gaze guide where a small model spends its attention.