Duck or rabbit? EMNLP 2026 paper probes whether MLLMs report bistable images like humans
anilkseth · x · 2026-10-22
An EMNLP 2026 Main paper led by Sussex PhD student rtakatsky, featured by Anil Seth, studies how multimodal LLMs report bistable images like the duck-rabbit illusion. Humans flip between interpretations under visual cues or expectations, typically seeing one at a time; the paper tests whether and how MLLMs show human-like reporting patterns.
More from Multimodal
- Kokon 0.10 reuses Qwen-Image 2.1's text encoder to write prompts, runs on 8GB cards — Ok-Director4174 · 2026-10-22
- Training-free 20B model Viveki generates real-time video heatmaps under 150ms — not-ekalabya · 2026-10-22
- Speechify launches Simba 3.5: one streaming-native TTS model across 48 languages — baseten · 2026-10-22
- Designer Khyati Trehan uses Gemini 4 Argon 3D for personal illustration work — alexanderchen · 2026-10-22
- Full 3:26 music video generated locally on a single AMD RX 7800 XT with open-source tools — wilppuse · 2026-10-22
- Fictional Lumea Serum ad shows how convincing AI-generated commercials have become — Amaan3024 · 2026-10-22