Original title: IchthyoNoma: Nomenclature and Context Sensitivity of Zero-Shot Biological Vision--Language Models for Bangladeshi Freshwater Fish Recognition
Nazim-E-Alam, Tarek Rahman, Md Kishor Morol · September 2026
Can artificial intelligence recognize fish just by matching pictures to written names? Vision-language models try to do this without any extra training. Researchers tested how well models like generic CLIPA common AI model that learns to match images with text descriptions to recognize objects., BioCLIPA version of CLIP adapted for biological images, like plants and animals, to improve identification., and BioCLIP2 identify freshwater fish from Bangladesh. These models match images to text descriptions. The specialized BioCLIP2 did much better than generic CLIP when using English common names or scientific names. However, when tested with Bengali names, BioCLIP2 fell to near-chance levels. A multilingual model did slightly better with Bengali, but also fell back to near chance with bare names. Testing image changes like blur or masking showed modest accuracy drops, large white-mask artifacts, and big differences between species. What looks like visual knowledge actually depends heavily on language, prompts, and context.
You're at a birdwatching event, and someone describes a bird using words from a language you don't speak well. Even though you see the bird clearly, you can't match the description to the bird in front of you. Here, the AI models are like birdwatchers who are experts in English names but clueless with Bengali ones, so their accuracy drops sharply when the language changes.
This matters for scientists and local communities in Bangladesh who use AI tools to identify fish species. If accuracy drops because of language mismatches, it could lead to errors in biodiversity conservation or local fish management.
The abstract doesn't state explicit limitations, but the work is confined to seven freshwater-fish categories from two Bangladeshi sources and tests only a few VLMs. It does not address other species, languages, or real-world deployment.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
IchthyoNoma: Nomenclature and Context Sensitivity of Zero-Shot Biological Vision--Language Models for Bangladeshi Freshwater Fish Recognition — Nazim-E-Alam, Tarek Rahman, Md Kishor Morol
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.