It is shown that mechanistic diagnosis can serve as a practical guide for targeted adaptation in underrepresented-language medical LLMs, and Targeted Low-Rank Adaptation (TLoRA) is proposed, restricted to the layer window where cross-lingual representations diverge, upstream of the output layers where the failure manifests.
Abstract
Large Language Models (LLMs) perform strongly in English medical tasks but degrade substantially in Arabic, a gap widely attributed to limited training data. We systematically investigate this assumption via tuned lens probing and causal activation patching, and find that Arabic medical knowledge is present in intermediate model representations but fails to surface at the output. This mechanistic insight motivates a targeted adaptation strategy: rather than fine-tuning the full network, we propose Targeted Low-Rank Adaptation (TLoRA), restricted to the layer window where cross-lingual representations diverge, upstream of the output layers where the failure manifests. We evaluate TLoRA on multiple-choice medical QA, where our approach outperforms full-network LoRA, zero-shot, and few-shot baselines. We further evaluate it on short-answer generation and multi-turn clinical dialogue, where it performs competitively without the need for task-specific finetuning. We additionally introduce AraClinicDialog, a clinician-constructed Arabic medical dialogue benchmark in MSA with validated variants across four Arabic dialects. Together, these contributions demonstrate that mechanistic diagnosis can serve as a practical guide for targeted adaptation in underrepresented-language medical LLMs.
Arabic clinical NLP systems often receive short, vague, or incomplete questions, which yields weak downstream answers even with strong encoders. We address this bottleneck by making question quality a first-class and measurable objective. Using domain-adaptive (continued) pretraining with a masked-language objective (DAPT-MLM) on AHQAD (~ 808k Arabic health Q–A pairs), we adapt two widely used backbones—AraBERT and the generator variant of AraELECTRA—to the lexical, syntactic, and discourse patterns of well-formed medical questions. Evaluation is aligned with the learning signal: we report cross-entropy and perplexity only at masked tokens, top-k accuracy restricted to masked spans, and lexical-diversity measures to discourage formulaic phrasing. A length-controlled test design (Short/Long/Very Long) isolates modeling gains from verbosity. Results show consistent intrinsic improvements for the domain-adapted models; AraBERT-MLM is best overall (macro Top-5 = 0.8392, lowest CE/PPL), outperforming AraBERT (orig.) by + 6.0 pp Top-5 and AraELECTRA (orig.) by + 17.2 pp. A 200-item human study (clinician + linguist) corroborates these gains (mean ± 95% CI: Clarity 4.12 ± 0.18, Fluency 3.68 ± 0.22, Semantic Fidelity 3.15 ± 0.25, Usefulness 3.42 ± 0.21; substantial agreement, κ ≈ 0.77) and highlights residual semantic drifts that inform simple, slot-constrained decoding fixes. Overall, the proposed reformulation module produces more natural and clinically relevant Arabic questions and can be plugged into Arabic clinical QA pipelines as a measurable, tunable front-end.
Walid Ounachad, M. Khenchouch, Imad Zeroual et al.· Language Resources and Evalu...· 0 citations
Developing artificial intelligence capable of clinical language comprehension and reliable diagnostic reasoning has remained a core challenge in biomedical engineering. While Large Language Models (LLMs) demonstrate significant potential in general natural language processing tasks, their direct application in the medical domain is severely constrained by parametric hallucinations and data silos. This paper introduces an end-to-end, resource-efficient, multilingual speech-driven Question-Answering (QA) framework optimized for localized clinical support. To accommodate deployment on consumer-grade execution environments, we implement Parameter-Efficient Fine-Tuning (PEFT) using Low-Rank Adaptation (LoRA) and 4-bit Quantized LoRA (QLoRA) configurations across open-source 3B and 7B parameter architectures. Human preference alignment is enforced via a stateful Reinforcement Learning with Human Feedback (RLHF) loop applying Proximal Policy Optimization (PPO). Crucially, to mitigate the vulnerabilities of passive information retrieval, we introduce an Active Validation Loop powered by Corrective Retrieval-Augmented Generation (CRAG). This validation engine is decoupled from the model harness using the Model Context Protocol (MCP), standardizing asynchronous lookups across dense vector repositories, clinical guidelines, and real-time electronic health registries.
Misha Patel· International Journal of Sci...· 0 citations
Medical visual question answering (VQA) is a crucial task in clinical AI, yet its evaluation has so far centered almost exclusively on English, limiting its relevance to linguistically diverse patients and clinicians. Recent multilingual medical VQA benchmarks show that large vision-language models (LVLMs) degrade in non-English languages, but lack a fine-grained analysis of how cross-lingual variation affects the distinct capabilities that medical VQA requires. To this end, we construct a multilingual medical VQA benchmark over eight languages, organized into four representative scenarios that isolate the core capabilities medical VQA requires. Evaluating five open- and closed-source LVLMs, we find that cross-lingual degradation is not uniform but highly scenario-dependent. We therefore propose MedVL-XLRepE, a training-free scenario-aware representation engineering method, leveraging LVLMs'superior English medical VQA capability to steer non-English representations toward their English counterparts at inference time. Across three LVLMs and eight languages, MedVL-XLRepE consistently mitigates cross-lingual degradation, with gains of up to 6.33\%.
Jingbo Wang, Sendong Zhao, Haochun Wang et al.· 0 citations
This study investigates the English-to-Turkish clinical translation performance of three large language models (Aya-23, Seed-X-Instruct, Gemma-3) using zero/few-shot prompting and QLoRA fine-tuning. Our three-stage setup evaluates: (1) general translation after FLORES-devtest training, (2) medical domain adaptation on i2b2 via a mixed dataset, and (3) generalization capacity on the independent FLORES-test set after mixed-data fine-tuning, using SacreBLEU, chrF, and COMET. Results show that Seed-X-Instruct achieves the strongest fine-tuned performance, reaching 32.70 BLEU, 64.34 chrF, and 90.65 COMET on the general FLORES test set, and 22.40 BLEU, 54.61 chrF, and 79.70 COMET on the medical i2b2 set. After mixed-data training, Seed-X improves from 11.48 to 32.39 BLEU on the unseen FLORES test set, whereas Aya-23 drops from 19.23 to 10.83 and Gemma-3 largely preserves its strong zero-shot baseline. These findings indicate that QLoRA effects are model-dependent and that domain adaptation does not yield the same generalization behavior across models.
Muhammetalp Erdem, Ş. Bora, Oguz Dikenelli· Signal Processing and Commun...· 0 citations
An engineering-oriented, end-to-end roadmap that structures the full lifecycle of clinical language model systems—from model design and domain adaptation to optimization and real-world evaluation is introduced.
Distractor selection rates in multiple-choice questions (MCQs) provide a behavioral proxy for distractor plausibility, yet current evaluation methods are reactive and resource-intensive. We investigated whether domain-specific fine-tuning of pre-trained embedding models can predict distractor selection rates from textual features before test administration. Using 6000 medical MCQs across eight clinical disciplines, we evaluated five general-purpose and five medical domain-specific embedding models under a unified 5-fold cross-validation protocol. Fine-tuning produced substantial improvements across both model categories: among medical models, SapBERT improved from
r
= 0.403 to
r
= 0.644 (+59.9%), while BGE-large improved from r = 0.467 to
r
= 0.626 (+34.0%) within the general group. Compared with lexical baselines where TF-IDF with string overlap features achieved the best performance (
r
= 0.546), the proposed transfer learning with fine-tuned contextual models showed meaningful improvement. Meanwhile, compact models also performed competitively: MiniLM (22 M parameters) reached
r
= 0.627 and MedEmbed-small (33 M) reached
r
= 0.629. These results establish the technical feasibility of text-based distractor selection rate prediction and characterize the performance landscape across model categories. This article offers a methodological investigation of a plausibility proxy, with potential application scenarios requiring future validation.
Zhehan Jiang, Tianpeng Zheng, Jiayi Liu et al.· npj Digital Medicine· 0 citations