When Chain-of-Thought Backfires: Evaluating Prompt Sensitivity in Medical Language Models

Small open-source language models are gaining attention for healthcare applications in low-resource settings where cloud infrastructure and GPU hardware may be unavailable. However, the reliability of these models under different phrasings of the same clinical question remains poorly understood. We evaluate five open-source models (Gemma 2 2B, Phi-3 Mini 3.8B, Llama 3.2 3B, Mistral 7B, and Meditron-7B, a domain-pretrained model without instruction tuning) across three clinical question answering datasets (MedQA, MedMCQA, and PubMedQA) using five prompt styles: original, formal, simplified, roleplay, and direct. Model behavior is evaluated using consistency scores, accuracy, and instruction-following failure rates. All experiments were conducted locally on consumer CPU hardware without fine-tuning. Consistency and accuracy were largely independent across models. Gemma 2 achieved the highest consistency (0.845-0.888) but the lowest accuracy (33.0-43.5%), while Llama 3.2 showed moderate consistency (0.774-0.807) alongside the highest accuracy (49.0-65.0%). Roleplay prompts consistently reduced accuracy across all models, with Phi-3 Mini dropping 21.5 percentage points on MedQA. Meditron-7B exhibited near-complete instruction-following failure on PubMedQA (99.0% UNKNOWN rate), indicating that domain pretraining alone is insufficient for structured clinical question answering. These findings show that high consistency does not imply correctness: models can be reliably wrong, a dangerous failure mode in clinical AI. Llama 3.2 demonstrated the strongest balance of accuracy and reliability for low-resource deployment. Safe clinical AI requires joint evaluation of consistency, accuracy, and instruction adherence.

Incorporating large language models (LLMs) in medical question answering demands more than high average accuracy: a model that returns substantively different answers each time it is queried is not a reliable medical tool. Online health communities such as Reddit have become a primary source of medical information for millions of users, yet they remain highly susceptible to misinformation; deploying LLMs as assistants in these settings amplifies the need for output consistency alongside correctness. We present a practical, open-source evaluation framework for assessing small, locally-deployable open-weight LLMs on medical question answering, treating reproducibility as a first-class metric alongside lexical and semantic accuracy. Our pipeline computes eight quality metrics, including BERTScore, ROUGE-L, and an LLM-as-judge rubric, together with two within-model reproducibility metrics derived from repeated inference (N=10 runs per question). Evaluating three models (Llama 3.1 8B, Gemma 3 12B, MedGemma 1.5 4B) on 50 MedQuAD questions (N=1,500 total responses) reveals that despite low-temperature generation (T=0.2), self-agreement across runs reaches at most 0.20, while 87-97% of all outputs per model are unique -- a safety gap that single-pass benchmarks entirely miss. The clinically fine-tuned MedGemma 1.5 4B underperforms the larger general-purpose models on both quality and reproducibility; however, because MedGemma is also the smallest model, this comparison confounds domain fine-tuning with model scale. We describe the methodology in sufficient detail for practitioners to replicate or extend the evaluation for their own model-selection workflows. All code and data pipelines are available at https://github.com/aviad-buskila/llm_medical_reproducibility.

Practitioners deploying small open-weight large language models (LLMs) for medical question answering face a recurring design choice: invest in a domain-fine-tuned model, or keep a general-purpose model and inject domain knowledge at inference time via retrieval-augmented generation (RAG). We isolate this trade-off by holding model size, prompt template, decoding temperature, retrieval pipeline, and evaluation protocol fixed, and varying only (i) whether the model has been domain-adapted (Gemma 3 4B vs. MedGemma 4B, both 4-bit quantized and served via Ollama) and (ii) whether retrieved passages from a medical knowledge corpus are inserted into the prompt. We evaluate all four cells of this 2x2 design on the full MedQA-USMLE 4-option test split (1,273 questions) with three repetitions per question (15,276 LLM calls). Domain fine-tuning yields a +6.8 percentage-point gain in majority-vote accuracy over the general 4B baseline (53.3% vs. 46.4%, McNemar p < 10^-4). RAG over MedMCQA explanations does not produce a statistically significant gain in either model, and in the domain-tuned model the point estimate is slightly negative (-1.9 pp, p = 0.16). At this scale and on this benchmark, domain knowledge encoded in weights dominates domain knowledge supplied in context. We release the full experiment code and JSONL traces to support replication.