Glonce

Study: country name in prompt shifts Qwen3 8B kidney disease screening

researchThe Glonce editorsarXiv🔥57Glonce rating57Impact40Novelty50Relevance95Surprise45Players5Credibility75

A new arXiv preprint asks a narrow question: if a patient's clinical data stays exactly the same and only the country named in the prompt changes, does an LLM's chronic kidney disease (CKD) screening answer change too? According to the study, it does.

The authors used a community-based CKD dataset from Bangladesh with 284 participants, 112 with CKD and 172 without. They varied the country attribute across 25 global contexts while holding all clinical inputs constant, and tested two open-weight models, Qwen3 8B and Mistral 7B. Experiments ran on a common-correct subset of 156 participants, meaning those for whom both models produced correct baseline predictions.

The results were uneven. Non-CKD predictions were highly stable across geographic contexts, and both models gave consistent non-CKD predictions in nearly all settings. CKD-positive predictions were considerably more sensitive. For Qwen3, the CKD prediction figure dropped from a baseline of 65 to as low as 50 across country contexts. The abstract does not state the unit of those figures. Mistral showed smaller but non-negligible variation.

The authors also looked at how the 25 country contexts clustered in each model's behaviour. The groupings did not follow known epidemiological or healthcare similarities. Instead, they appear to reflect how country tokens are internally represented within each model.

The authors conclude that geographic tokens can systematically shift clinical screening outcomes despite identical patient inputs, with implications for deploying LLM-based decision support in diverse healthcare settings. The setup is zero-shot: the models were prompted directly, with no task-specific training described.

Key facts

  • Only the country named in the prompt was changed, across 25 country contexts, with all clinical inputs held constant.
  • Two open-weight models were tested, Qwen3 8B and Mistral 7B, on a Bangladeshi community CKD dataset of 284 participants (112 CKD, 172 non-CKD); experiments used a common-correct subset of 156.
  • Non-CKD predictions stayed stable in nearly all settings, while CKD-positive predictions were considerably more sensitive to country.
  • For Qwen3, the CKD prediction figure fell from a baseline of 65 to as low as 50 depending on country; Mistral varied less but not negligibly.
  • Country groupings did not match known epidemiological or healthcare similarities; they appear to follow how each model represents country tokens.

Why it matters

Screening tools are only useful if the answer depends on the patient. This study isolates one irrelevant variable, the country named in the prompt, and finds that it can still move the output. The effect fell on CKD-positive predictions, the cases a screening tool exists to catch. For Qwen3 the CKD prediction figure went from a baseline of 65 to as low as 50. The authors also report that the pattern of which countries behave alike does not track epidemiology or healthcare systems, so the shifts are not a sensible adjustment for local disease burden. They appear to come from how the model represents country tokens.

Who it affects

Anyone building or evaluating LLM-based clinical decision support, especially for use across different countries and healthcare settings, which the authors name as the area of concern. Teams working with open-weight models such as Qwen3 8B and Mistral 7B have direct evidence on two models. Researchers studying prompt sensitivity and bias in medical LLMs also get a concrete test design.

How to use it

The method is simple to copy: take a fixed set of patient records, keep every clinical input constant, vary only the country attribute across many contexts, and compare predictions with the baseline. The study restricted this to participants whom both models classified correctly at baseline, so any change can be attributed to the country token rather than to existing errors. The abstract does not state any mitigation or fix for the geographic sensitivity.

How solid is it

This is a preprint abstract describing a small study: 284 participants from one Bangladeshi community dataset, two 7B to 8B open-weight models, and a 156-person subset used for the experiments. The headline Qwen3 numbers (65 to 50) come without a stated unit. The abstract does not say which countries produced the lowest Qwen3 figure or which countries clustered together, and it gives no specific Mistral numbers. The explanation about internal representation of country tokens is stated as what the groupings appear to reflect, not as a demonstrated mechanism.

Risks and caveats

The findings cover only Qwen3 8B and Mistral 7B, and no result is reported for other models. The study is zero-shot only, so it makes no claim about fine-tuned or deployed clinical systems. The abstract does not say whether the shift was toward higher or lower disease risk for any particular country beyond the drop in CKD predictions for Qwen3. Because the dataset is a single community sample, the results should be read as a warning sign about prompt sensitivity, not as a measure of how any real-world tool would behave.

“geographic tokens can systematically shift clinical screening outcomes despite identical patient inputs”

— From the paper's abstract on arXiv