Unverified50% confidenceFactExact time
The subliminal learning phenomenon was previously observed mainly in text data, and the new study confirms that it also exists in the image selection step.
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen text instructions precede image input, the Transformer's cross-attention mechanism causes the model to have already formed clear query vectors when encoding the image, preferentially activating visual regions semantically related to the target.70% similarUnverifiedPrompt adherence and photorealism do not always improve in the same direction; checkpoints fine-tuned on large volumes of real photographs excel in skin texture and lighting, but may have weaker comprehension of complex prompts than base models.68% similarUnverifiedContrastive learning naturally encourages models to capture the most discriminative semantic features in an image, rather than fine-grained spatial and quantitative information.68% similarUnverifiedFor vision-language models, image resolution affects the recognition accuracy of chart and document details, and the number of visual tokens directly affects the remaining usable context window.68% similarUnverifiedThe paper points out that learned expression encodings suffer from severe view bias, i.e., the encoding results unintentionally absorb information about camera viewpoint, lighting, and head orientation.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1079339API
curl https://kongchang.com/api/v1/knowledge/claims/1079339MCP
get_claim(id=1079339)