Prompting strategies beyond labelled examples for untrained vision–language models recognising eosinophils in oesophageal biopsies: an exploratory study

Background. In a companion study, untrained vision–language models (VLMs) recognised eosinophils in oesophageal biopsy crops reliably only with labelled image examples. Whether other prompting strategies do better, or the results reflect memorised public images, is unknown. Methods. On the companion study's 1 200 crops (17 slides, 15 patients; reference by one pathologist, the author), the pre-specified primary comparison was balanced accuracy with argued answers (features and arguments first) versus direct answers in three VLMs (exact McNemar, Holm-adjusted). Exploratory protocol-fixed phases examined separate feature calls, four image-prompt variants in six VLMs, a transposed query, alternative example sets, 600 external colorectal crops and two local open-weight models. 95% CIs: slide bootstrap (region bootstrap externally). Results. Arguing did not improve balanced accuracy: −0.002 (−0.017 to 0.014) in Qwen3.7 Flash, −0.003 (−0.029 to 0.016) in Gemini 3.5 Flash-Lite and −0.020 (−0.033 to −0.006; Holm-adjusted McNemar p = 0.071, not significant) in Claude Haiku 5.5, at 1.2–1.9 times the cost, shifting the operating point towards eosinophil. Claude Haiku 5.5's labels computed from its reported features beat its direct answer (+0.023, 0.007 to 0.044), and a separate feature call added 0.083 (0.064 to 0.103). Grey squares replacing the example images cost five models 0.042–0.107; no variant improved the four stronger models, whereas curated or retrieved examples raised the weakest, GPT-6 Luna, by 0.051 (0.032 to 0.074; 0.027 to 0.074). On external crops the four stronger models reached 0.88–0.91 (0.83–0.87 against eosin-rich look-alikes added post hoc) and a logistic regression on foundation-model embeddings 0.968 (0.955–0.982). A local 4-billion-parameter model was 0.018 (−0.046 to 0.009) below the best paid model. Conclusions. Beyond labelled examples, no prompting strategy improved recognition in the four stronger models. A transposed query did not lower accuracy in the two stronger models tested, which argues against memorisation of these exact images. These exploratory results are hypothesis-generating. Companion preprint: doi:10.5281/zenodo.24061169. Files: main text and supplementary material.