Accurate but Unreliable?
While Large Language Models (LLMs) are increasingly used in healthcare, most benchmarks focus solely on average accuracy. This co-authored MSc team study tackles a more critical, often ignored dimension: consistency.
If a clinical question is posed repeatedly, does the model give the same answer every time — and does that hold up across languages? For the roughly 600 million Hindi speakers worldwide, this “self-agreement” gap is a major barrier to safe telehealth adoption.
The Methodology: 6,000 Inferences, Two Languages
The team evaluated three medical LLMs — Aloe-Beta-8B, Meditron3-7B, and Gemini 2.5 Flash-Lite — on 100 stratified USMLE-style multiple-choice questions, each run 10 times in both English and Hindi: 6,000 total inferences.
Statistical Framework
Each of the 10 runs was treated as an independent rater, enabling:
- Fleiss’ Kappa — measuring how consistently a model agrees with itself across runs.
- Krippendorff’s Alpha — a chance-corrected reliability measure.
- Bootstrapped 95% confidence intervals and Bonferroni-corrected Mann-Whitney U tests to confirm differences between languages were statistically significant, not noise.
The Result: A 23.3-Point Reliability Gap
The 8B-parameter specialist models showed a 23.3 kappa-point drop in self-agreement moving from English to Hindi — meaning the same clinical question, asked repeatedly in Hindi, was substantially more likely to get a different answer each time than the identical question asked in English. Average accuracy alone would never surface this: a model can look competent on a single pass while being meaningfully less trustworthy under repeated use.
Impact: Making AI Stable in Practice
This shifts the conversation from “is the AI smart?” to “is the AI stable?” — an essential distinction for deploying LLM-powered health tools to non-English-speaking populations, where a single wrong answer carries real clinical risk.