Reproduce “SPOOK: Sparse Pixel Optimization for Off-Manifold Knowledge in Vision Encoders” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ankit Gupta, Akshat Pandya, Christoph Adami, Emily Dolson

The authors introduce SPOOK, which starts with a white canvas and retains random single-channel edits only when they increase similarity to a chosen target. In 1,000-target retrieval, it achieves 94.1% top-one success with OpenCLIP and 85.8% with DINOv3, changing under 1% of pixel-channels. Results concern the fixed encoder and untransformed inputs; transfer is weak and transformations reduce performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.