@misc{indiciae2d92fba6c758,
  title = {Decodable but Misrouted: Sparse Features Uncover a Readout Gap in Vision-Language Models for Harmful Meme Detection},
  author = {Girish A. Koushik and Diptesh Kanojia and Helen Treharne},
  year = {2026},
  url = {https://arxiv.org/abs/2609.18860},
  note = {Source identifier: 2609.18860}
}