MedPerturbing LLMs: A Comparative Study of Toxicity, PromptTuning, and Jailbreaks in Medical QA
Bibliographic record
Abstract
Large Language Models (LLMs) are increasingly adopted across domains, including sensitive areas such as healthcare. However, their deployment raises significant safety concerns, particularly with respect to toxicity. In this paper, we evaluate the toxicity of widely used general-purpose LLMs in medical question–answering tasks. We investigate three complementary scenarios: (i) baseline querying, (ii) prompt guidelines designed to mitigate toxic outputs, and (iii) adversarial jailbreak prompting intended to elicit harmful content. To measure toxicity, we apply three established metrics to five LLMs ranging from 2B to 9B parameters, using MedPerturb, a dataset of medical questions systematically perturbed across gender, race, and age. Our results show that while carefully crafted guidelines can reduce toxic outputs and mitigate demographic biases, adversarial instructions are highly effective at bypassing safety mechanisms. Our evaluation reveals that all models exhibit limited resilience to jailbreak attacks, highlighting a critical vulnerability that restricts their safe deployment in clinical contexts. By answering three key questions—(1) what levels of toxicity these models exhibit in standard medical scenarios, (2) how far prompt tuning can reduce toxicity, and (3) how vulnerable they are to jailbreaks, our study provides a structured assessment of the risks and limitations of LLMs in healthcare, and shows the importance of establishing robust guidelines and protections to promote the safe deployment of LLMs in healthcare and to guard against harmful misuse.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.015 | 0.080 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.002 | 0.003 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".