Researchers have investigated whether multimodal large language models (MLLMs) exhibit human-like reporting behavior when presented with bistable images, such as the classic duck-rabbit illusion. Using the LLaVA family of models, the study explored how visual cues and linguistic priors influence model responses, finding that these factors systematically shift reports in ways consistent with human perception. The models also predominantly committed to a single interpretation, similar to humans, with internal computations involving competing image-token representations and distinct modulation pathways. AI
IMPACT Investigates how MLLMs process ambiguous visual information, potentially informing future model development for more nuanced perception.
RANK_REASON Academic paper detailing research on MLLM behavior with bistable images. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- duck-rabbit
- Gotit.pub
- Hugging Face
- Llava
- ScienceCast
- Visual Anagrams
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
