LLaVA-1.5
A family of open-source vision-language models (VLMs) built on the LLaMA architecture combined with a visual encoder, used for image-text understanding and multimodal dialogue research.
Timeline (last 90 days)
In the owl target experiment, the student model trained on data selected by a schemer that prefers owls exhibits a higher target-to-alternative likelihood ratio in image-free readout, and this directional shift appears on LLaVA-1.5, Qwen2.5-VL, Qwen3-VL, and Gemma 4.
In three models, the student selected with the owl target scores higher than all quantity-matched randomly reselected students; in LLaVA-1.5, it scores higher than 55 of 60 randomly reselected students.
The first-generation student can pass the preference to the second generation through descriptions that do not explicitly mention owls, and this intergenerational transmission is verified in both LLaVA-1.5 and Qwen3-VL.
Porting LLaVA-1.5's token IDs as-is to two models with disjoint vocabularies drops F1 straight to 0.00, whereas using tokenizer-derived readings keeps F1 stable between 0.80 and 0.92.
The context length of mainstream open-source multimodal models is typically between 4K and 32K tokens: LLaVA-1.5 is 4,096 tokens, Qwen-VL is 8,192 tokens, and InternVL2 extends to 32K.
LLaVA-1.5 and its follow-up work introduced two-layer MLPs and even cross-attention mechanisms to enhance cross-modal information fusion.
All Facts (6)
In the owl target experiment, the student model trained on data selected by a schemer that prefers owls exhibits a higher target-to-alternative likelihood ratio in image-free readout, and this directional shift appears on LLaVA-1.5, Qwen2.5-VL, Qwen3-VL, and Gemma 4.
50%UnverifiedIn three models, the student selected with the owl target scores higher than all quantity-matched randomly reselected students; in LLaVA-1.5, it scores higher than 55 of 60 randomly reselected students.
50%UnverifiedThe first-generation student can pass the preference to the second generation through descriptions that do not explicitly mention owls, and this intergenerational transmission is verified in both LLaVA-1.5 and Qwen3-VL.
50%UnverifiedPorting LLaVA-1.5's token IDs as-is to two models with disjoint vocabularies drops F1 straight to 0.00, whereas using tokenizer-derived readings keeps F1 stable between 0.80 and 0.92.
50%UnverifiedThe context length of mainstream open-source multimodal models is typically between 4K and 32K tokens: LLaVA-1.5 is 4,096 tokens, Qwen-VL is 8,192 tokens, and InternVL2 extends to 32K.
50%UnverifiedLLaVA-1.5 and its follow-up work introduced two-layer MLPs and even cross-attention mechanisms to enhance cross-modal information fusion.
50%