Field testing of vision and language models shows narrow specialization beats size and poses risks of fabricating categories

Remote field environments such as natural reserves and wildlife monitoring stations impose critical constraints on AI deployment, because monitoring cameras typically run on low-power edge devices without internet connectivity. This reality makes locally runnable embedded models, not giant models, the only practical benchmark for evaluation. In this context, a research team led by William Zou, Mayouka Siriboram, Xiao Yan, Ziki Liu, and Yi Ding tested, in a paper published on arXiv (preprint repository) with identifier 2609.11916, how general vision-language models ranging from 2 billion to 8 billion parameters perform on genuine classification tasks, by comparing four general models, Qwen-3-VL at 2 billion, 4 billion and 8 billion parameters, and Gemini-3 at 4 billion parameters, against the specialized BayClip model, which has only 300 million parameters.

The experiment subjected the models to a classification task covering 96 biological species, by comparing clean laboratory images from the iNaturalist platform with real-world shots captured by field monitoring cameras across six field-image datasets.The ultra-small specialized model outperformed all general models by a wide margin.BayClip surpassed all vision-language models by 33.2 to 59.2 percentage points on an expanded sample of 200 images per model, even though its size represents only a tiny fraction of the parameters. This result clearly shows that the gap in classification accuracy is not linked to model architectural scale but primarily to the specialized nature of the training data, which acts as a core driver of knowledge efficiency.

Although all models exceeded the random-guess baseline, the tests revealed a sharp performance drop when moving from clean images to field shots, with a decline ranging from 9.6 to 26.6 percentage points across the various classification levels.The drop in field accuracy reflects the difficulty of reading and the noise in the images, not a failure of fine discrimination.BayClip itself showed a decrease of 18.0 percentage points, a gap that is not statistically distinguishable from the 22.3-point drop recorded by the best general model. This means that the loss of effectiveness between the clean test environment and real-world field conditions is not a flaw specific to general models, but a property imposed by lighting variations, capture angles, and visual noise in field sensors.

When the models were tested with open-ended prompting without restricting the options, a severe operational issue emerged: the proportion of responses that contained syntactically and linguistically correct but scientifically nonexistent and entirely fabricated entity names ranged from 5.9 % to 9.6 %. The ranking of models in hallucination rates reproduced precisely across two independent evaluation sets, confirming that the generation of fictitious entities is a stable structural behavior of unrestricted models and could corrupt scientific records if the models are left to operate without strict verification rules.

This measurement carries direct operational implications for organizations and technical teams in the Gulf, Egypt, and the Arab Levant, especially as projects expand in natural reserves, biodiversity monitoring in desert and marine environments, and agricultural pest surveillance in areas lacking cloud coverage. It is no longer worthwhile to load edge devices with general models that consume processing power and battery life in the hope that an 8-billion-parameter size compensates for a lack of specialization.Investing in embedded, domain-specific models for biology yields higher accuracy with double the computational efficiency.Avoiding classification hallucinations requires abandoning open-ended prompts and using closed taxonomy lists that constrain outputs to the targeted local species. The results also confirm for field engineers that upgrading camera sensors and applying local image-enhancement techniques constitute half the battle, as poor field shots can cost roughly a quarter of a model’s accuracy regardless of its overall quality.