Diagnosis-Aware Prompt Strategy

• Diagnosis-aware prompt strategy is a design principle that embeds clinical evidence and structured reasoning into prompts to guide diagnostic decisions.
• It leverages methods including concept tokens, chain-of-thought, and knowledge-injected templates across imaging, NLP, and multimodal systems to enhance diagnostic performance.
• The approach improves model interpretability and robustness by aligning computational processes with clinical decision-making, yielding measurable gains on benchmark tasks.

Searching arXiv for recent and relevant papers to ground the article. Diagnosis-aware prompt strategy is a family of methods in which prompts encode clinically meaningful concepts, diagnostic reasoning patterns, or structured disease evidence so that model computation is organized around diagnosis rather than generic task conditioning. Across recent work, this idea appears as concept prompts inside frozen vision backbones, chain-of-thought prompts that force clinician-like reasoning, knowledge-injected masked-language templates for clinical text, prompt-generated seeds for diagnosis extraction, graph prompts for multimodal neuroimaging, and multi-agent consultation flows for interactive diagnosis (Dong et al., 4 Oct 2025, Savage et al., 2023, Zheng, 2024, Chuang et al., 2023, Peng et al., 2023, Sanghvi et al., 2 Jun 2026).

1. Definition and conceptual scope

In medical AI, “diagnosis-aware” does not denote a single prompt template. It denotes a design principle: prompts are made to reflect the structure of clinical decision making, the semantics of disease concepts, or the evidentiary constraints of diagnosis. In CoPA, diagnosis-aware prompting means using concept-structured, task-specific prompts so that every layer becomes aware of clinically relevant concepts and the final diagnosis is produced by aggregating and weighting those concept representations (Dong et al., 4 Oct 2025). In LLM-based clinical reasoning, the same phrase is operationalized as prompts that explicitly ask the model to reason through differential diagnosis formation, intuitive reasoning by pattern/association, analytic reasoning via pathophysiology, or Bayesian inference (Savage et al., 2023). In prompt learning for clinical text, it means placing disease-related decisions at [MASK] positions and injecting knowledge-graph evidence into soft tokens so that BERT is guided by diagnostic semantics rather than by generic classification cues (Zheng, 2024).

This diversity can be organized by the object being prompted. Some methods prompt internal representations, some prompt reasoning style, and some prompt system workflow.

Setting	Prompt carrier	Representative works
Medical imaging and VLMs	Concept tokens, soft prompts, graph prompts	CoPA, XCoOp, MMGPL
Clinical LLM diagnosis	CoT instructions, differential lists, few-shot rationales	DR-CoT, diagnostic reasoning prompts, MeDxAgent
Clinical NLP and report generation	[MASK] templates, soft tokens, diagnosis tokens	knowledge-enhanced BERT, PromptMRG
Structured extraction	Few-shot schema prompts	periodontal diagnosis extraction
Robustness optimization	prompt groups, stability objectives, reflective prompt revision	prompt group-aware training, stability-aware optimization, RPT

Taken together, these papers suggest that diagnosis-aware prompting is best understood as an interface between model adaptation and clinical epistemology: the prompt is designed not merely to elicit an answer, but to impose a clinically legible decomposition of evidence, concepts, and decision rules.

2. Concept-structured prompting in imaging and multimodal diagnosis

The most explicit imaging formulation is CoPA, a concept bottleneck framework for dermoscopy and clinical skin images. Given triplets $\mathcal{D}=\{(x,c,y)\}$, CoPA predicts human concepts $c=\{c_1,\dots,c_N\}$ from image $x$, then predicts disease label $y$ from those concepts. Its Concept-aware Embedding Generator extracts concept embeddings $\mathbf{z}_l^i$ from every transformer layer of a BiomedCLIP image encoder, and these embeddings are reused as prompt tokens by Concept Prompt Tuning in the next layer while the backbone remains frozen (Dong et al., 4 Oct 2025). The framework optimizes concept alignment and diagnosis jointly,

$$\mathcal{L}=\lambda \mathcal{L}_{con}+(1-\lambda)\mathcal{L}_{CE}(\hat y,y),$$

so prompts are trained to be simultaneously concept-accurate and diagnosis-useful. This multilayer prompting matters because many clinical concepts are low-level or local/multi-scale rather than final-layer global semantics. On PH$^2$, Derm7pt, and SkinCon, CoPA outperforms state-of-the-art methods for both concept prediction and disease prediction; on Derm7pt, for example, disease accuracy improves by about $2.1\%$ over the second-best baseline, and concept accuracy improves by $2.6\%$ over the second-best method on PH$^2$ (Dong et al., 4 Oct 2025).

A related but distinct line uses concept-guided soft prompts for explainable CAD. XCoOp aligns learnable prompts with clinical concept-driven prompts at token level and prompt level, and aligns both with global and local image features from CLIP (Bie et al., 2024). Its clinical prompts are built either from expert concept annotations, as in Derm7pt and SkinCon, or from GPT-4-elicited visual concepts for datasets without concept labels. The reported gains are diagnostic as well as interpretive: XCoOp improves AUC and accuracy across Derm7pt, SkinCon, Pneumonia, and IU X-Ray, while also reducing the distance between learned soft prompts and clinical prompts from 1.982 or 3.011 in prior methods to 1.104 (Bie et al., 2024). This suggests that diagnosis-aware prompting can be embedded directly into the geometry of the prompt space.

MMGPL extends the same principle to multimodal neuroimaging. It uses GPT-4 to generate disease-related concepts, computes semantic similarity between those concepts and image patches, down-weights irrelevant patches, constructs a concept-conditioned graph over tokens, and uses a GCN to produce graph-prompted embeddings for a frozen multimodal transformer (Peng et al., 2023). In ADNI and ABIDE, MMGPL achieves the best reported results among the compared prompt-learning and multimodal-large-model baselines; on ADNI-3CLS it reaches ACC $c=\{c_1,\dots,c_N\}$0 and AUC $c=\{c_1,\dots,c_N\}$1, improving over MaPLe and MetaT (Peng et al., 2023). Here, diagnosis awareness is not only semantic but structural: the prompt encodes which regions matter and how disease-relevant regions relate as a graph.

Prompt sensitivity itself has also become an object of training. In text-guided nuclei segmentation, semantically related prompts are organized into prompt groups sharing one mask, and training adds a quality-guided group regularization plus a logit-level consistency loss with stop-gradient (Wu et al., 6 Mar 2026). The method leaves inference unchanged yet reduces variance across low-, medium-, and high-quality prompts, and on six zero-shot cross-dataset tasks improves Dice by an average of 2.16 points (Wu et al., 6 Mar 2026). A plausible implication is that diagnosis-aware prompting in pathology increasingly requires not only semantic specificity but invariance to clinically harmless paraphrase.

3. Clinician-like reasoning prompts for LLMs

In LLM-based diagnosis, diagnosis-aware prompting usually means constraining the reasoning form rather than inserting concept tokens into a backbone. One influential formulation is “Diagnostic-Reasoning Chain-of-Thought” (DR-CoT), which makes the model summarize current evidence, formulate a ranked differential diagnosis, choose the next question to narrow the differential, and iterate until final diagnosis (Wu et al., 2023). On DDXPlus, two DR-CoT exemplars improve diagnostic accuracy by 15% relative to standard prompting, with an 18% gain in out-domain settings, and physician review judged DR-CoT questions more clinically critical than baseline prompts in most sampled dialogues (Wu et al., 2023). This is diagnosis-aware because the prompt fixes the intermediate artifacts of clinical reasoning—evidence summary, DDx, and discriminative question selection.

A second formulation decomposes clinician reasoning into explicit prompt modes. “Diagnostic Reasoning Prompts Reveal the Potential for LLM Interpretability in Medicine” defines four canonical prompt strategies: differential diagnosis formation, intuitive reasoning by pattern/association, analytic reasoning via pathophysiology, and Bayesian inference (Savage et al., 2023). On free-response MedQA diagnostic questions, GPT-4 maintains roughly the same diagnostic accuracy under traditional CoT and diagnosis-aware prompts: traditional CoT 76%, intuitive 77%, differential 78%, analytic 78%, Bayesian 72% on the reduced 497-question test set (Savage et al., 2023). By contrast, GPT-3.5 deteriorates under most structured reasoning prompts. The paper’s central claim is not that one reasoning mode universally maximizes accuracy, but that stronger models can be required to expose clinician-recognizable reasoning without sacrificing performance.

Clinical CoT has also been used as direct rationale supervision. The reasoning-aware diagnosis framework of Chen et al. generates clinical rationales with GPT-4, then uses them both for few-shot prompting and for distillation into smaller unimodal and multimodal student models (Kwon et al., 2023). The framework formalizes reasoning-aware diagnosis as

$c=\{c_1,\dots,c_N\}$2

rather than direct label prediction from patient description alone. On ADNI, GPT-4 with 2-shot Clinical CoT reaches 68.4 accuracy versus 62.6 for 5-shot standard prompting and 59.6 for zero-shot prompting, and the paper evaluates rationales using clinician-defined criteria of consistency, correctness, specificity, helpfulness, and human-likeness (Kwon et al., 2023). This line treats diagnosis-aware prompts as a mechanism for eliciting clinically structured latent reasoning and as a source of training data for smaller deployable systems.

The interactive endpoint of this trend is multi-agent diagnosis. MeDxAgent models diagnosis as consultation rather than one-shot classification: it collects demographics, maintains summarized dialogue, produces top-3 candidate diagnoses with rationales and confidences, uses evidence-gap analysis, and switches to differential-questioning after enough information has accumulated (Sanghvi et al., 2 Jun 2026). On MeDxBench, a benchmark of 4,421 cases across 20 specialties, MeDxAgent achieves a 10.3% accuracy gain over baseline and closes 52.3% of the gap to a full-information oracle (Sanghvi et al., 2 Jun 2026). Critically, the paper shows that diagnosis-guided questioning is timing-sensitive: switching to differential questioning at turn 2 is harmful, while switching at turn 10 is beneficial. Diagnosis-aware prompting in this setting is therefore not only about prompt content but also about prompt scheduling within a consultation flow.

4. Knowledge injection, prompt tuning, and diagnosis-conditioned generation

Another major formulation turns diagnosis into a prompt-conditioned classification or generation problem by explicitly injecting disease knowledge. A knowledge-enhanced BERT framework converts disease diagnosis into masked language modeling with mixed discrete and soft prompts, where knowledge-graph reasoning paths are encoded and injected into soft tokens (Zheng, 2024). The task is reframed as

$c=\{c_1,\dots,c_N\}$3

and the soft token is updated by averaging with related vectors,

$c=\{c_1,\dots,c_N\}$4

Reasoning paths such as disease-to-symptom or symptom-to-complication chains are verbalized, encoded with BERT, and used to make the prompt explicitly sensitive to diagnostic knowledge (Zheng, 2024). The paper reports F1 improvements of 2.4% on CHIP-CTC, 3.1% on IMCS-V2-NER, and 4.2% on KUAKE-QTR in the abstract, and its ablation shows a drop from F1 0.91 to 0.89 when knowledge is removed (Zheng, 2024).

PromptMRG uses the same principle for report generation: a disease classification branch predicts four-way status labels—blank, positive, negative, uncertain—for each disease, and these labels are converted into diagnosis-driven prompt tokens such as [BLA], [POS], [NEG], and [UNC] that condition a report decoder (Jin et al., 2023). The model combines image encoding, CLIP-based cross-modal feature enhancement, a disease classification branch, and a BERT-base decoder. The overall loss is

$c=\{c_1,\dots,c_N\}$5

where the self-adaptive disease-balanced term uses logit adjustment based on disease learning status. On MIMIC-CXR, PromptMRG reaches CE F1 $c=\{c_1,\dots,c_N\}$6 versus $c=\{c_1,\dots,c_N\}$7 for RGRG, and ablation shows that adding diagnosis-driven prompts alone raises CE F1 from $c=\{c_1,\dots,c_N\}$8 to $c=\{c_1,\dots,c_N\}$9 (Jin et al., 2023). Here diagnosis-aware prompting is literally a discrete, structured interface from classifier to generator.

In medical LVLMs, prompt engineering has also been used to control hallucination and minority-pathology failure. A prompting study on LLaVA-Med introduces two strategies: provide a detailed explanation of the queried pathology, and provide the judgment of a weak learner trained for a specific error trade-off (Guo et al., 2024). The first strategy improves diagnostic F1 on several chest pathologies by supplying radiographic criteria in the prompt; the second textualizes the weak learner’s probability so that the LVLM can be nudged away from false positives or false negatives depending on task design. The paper reports the highest diagnostic F1 increase as 0.27 and Recall improvement of approximately 0.07 on POPE metrics (Guo et al., 2024). Diagnosis-aware prompting in this case is neither concept-bottleneck nor CoT; it is explicit contextualization of the pathology definition and of auxiliary diagnostic evidence.

A more extraction-oriented variant appears in periodontal diagnosis mining from dental records. GPT-J few-shot prompts map diagnosis-like sentences to a structured Stage/Grade/Extent triple, and the resulting seeds train a RoBERTa NER system (Chuang et al., 2023). The best direct few-shot GPT-J setting reaches F1 about 0.72 with a lower negative ratio and larger number of examples, while downstream RoBERTa models trained on GPT-J seeds achieve F1 in the 0.92–0.97 range across configurations (Chuang et al., 2023). This work shows that diagnosis-aware prompting can define a task schema—here a structured diagnosis tuple—even when the downstream model is not itself prompted at inference.

5. Explainability, faithfulness, and evaluation criteria

A recurring motivation for diagnosis-aware prompting is that prompts should not merely improve metrics but also expose a clinically intelligible decision path. CoPA is exemplary here. Its decision pathway is image $x$0 concepts $x$1 diagnosis, with concept confidence obtained from visual–text alignment scores and final diagnosis obtained through a gated aggregation module,

$x$2

The model supports concept heatmaps, concept confidences, gating weights, and full-pipeline visualization, and its test-time concept interventions on Derm7pt show that forcing correct concepts increases accuracy by 0.5–1.1% while flipping correct concepts to 0 decreases accuracy by 2.4–3.2% (Dong et al., 4 Oct 2025). The paper interprets this as evidence that the concept pathway is faithful rather than merely post hoc.

XCoOp pursues the same goal from the prompt side. Because learned prompts are explicitly aligned with clinical concept prompts, they admit textual explanation through proximity to those prompts and visual explanation through similarity maps between prompt embeddings and image regions (Bie et al., 2024). The paper’s concept-based knowledge intervention shows that replacing correct clinical concepts with random, general, or intervened knowledge degrades performance, which the authors treat as evidence that the model genuinely uses the concept information encoded in prompts (Bie et al., 2024).

Reasoning-based systems evaluate explanation quality more directly. The clinical rationale framework scores rationales for consistency, correctness, specificity, helpfulness, and human-likeness (Kwon et al., 2023). Diagnostic reasoning prompts for MedQA emphasize procedural interpretability rather than formal faithfulness metrics, but the reported result—that GPT-4 can follow clinician-like reasoning styles without sacrificing accuracy—was presented as a way for clinicians to inspect and judge the model’s logic (Savage et al., 2023). MeDxAgent similarly forces every diagnosis agent to output disease, confidence, and brief rationale in structured form, allowing downstream agents to use those rationales for evidence-gap analysis and targeted questioning (Sanghvi et al., 2 Jun 2026).

PromptMRG adds an important generative example: when generating disease terms such as “cardiomegaly,” decoder attention shifts toward the corresponding diagnosis-driven prompt tokens, showing that the report text is conditioned by explicit disease-status prompts rather than only by latent image features (Jin et al., 2023). This suggests that diagnosis-aware prompting often turns explainability into a systems property: the prompt object, whether concept token, rationale, or status token, is designed to be inspectable by construction.

6. Robustness, limitations, and emerging design principles

Diagnosis-aware prompting does not automatically guarantee robustness. Histopathology prompt engineering shows that prompt wording remains consequential even in highly specialized VLMs. In zero-shot digestive pathology, a systematic ablation across domain specificity, anatomical precision, instructional framing, and output constraints found that prompt engineering significantly impacts diagnostic accuracy, that CONCH performs best with precise anatomical references, and that performance consistently degrades when anatomical precision is reduced (Sharma et al., 30 Apr 2025). This result is closely aligned with diagnosis-aware prompting: clinical specificity helps when it matches the model’s training distribution, but verbose or misaligned instructions can still degrade performance.

Prompt stability has therefore become an explicit objective. “Stability-Aware Prompt Optimization for Clinical Data Abstraction” defines prompt sensitivity via flip rates under paraphrased prompts and shows that higher accuracy does not guarantee prompt stability (Kolbeinsson et al., 29 Jan 2026). Across MedAlign applicability/correctness and MS subtype abstraction, a dual-objective prompt optimization loop jointly targets accuracy and stability, often reducing flip rates substantially, sometimes at modest accuracy cost (Kolbeinsson et al., 29 Jan 2026). The paper’s central warning is that models can appear well-calibrated yet remain fragile to paraphrases. For diagnosis-aware systems, this is consequential because prompt wording is rarely static in deployment.

Reflective Prompt Tuning pushes this idea further by making prompt optimization itself diagnosis-driven. An optimizer LLM must call a diagnostic function that evaluates the target model over an optimization set, summarizes recurring failure modes into a structured diagnostic report, and returns it together with calibration signals (Bayat et al., 20 May 2026). The optimizer then revises the prompt using both the current report and a memory of prior reports. Across HotPotQA, LiveBench-Math, and Formula, RPT improves over initial prompts by up to 12.9 points and improves Brier score in confidence-aware selection (Bayat et al., 20 May 2026). Although these are not medical tasks, the method is directly relevant: it operationalizes prompt engineering as failure-mode diagnosis plus targeted revision, a pattern that is increasingly natural for clinical prompting where error taxonomies matter.

Several broader design principles recur across the literature. Clinically meaningful concepts should be chosen so that they cover both low-level and high-level evidence; prompts should shape internal computation rather than merely annotate outputs; frozen or lightly tuned backbones are often preferred in data-scarce medical settings; textual definitions or knowledge paths help anchor prompts to clinical semantics; and diagnosis-aware systems should be evaluated not only for task accuracy but also for faithfulness, calibration, stability under paraphrase, and behavior on rare or minority conditions (Dong et al., 4 Oct 2025, Zheng, 2024, Jin et al., 2023, Kolbeinsson et al., 29 Jan 2026). The literature also suggests several forward directions: extension to radiology, pathology, and other modalities with domain-specific concept sets; hierarchical or temporal concepts; richer instruction-style prompts; and prompt optimization procedures that explicitly treat robustness and calibration as coequal design objectives (Dong et al., 4 Oct 2025, Peng et al., 2023, Bayat et al., 20 May 2026).

In this sense, diagnosis-aware prompt strategy has evolved from a narrow prompting trick into a broader methodological program. The prompt becomes a clinically structured control surface through which concepts, differentials, knowledge graphs, retrieved evidence, confidence signals, and interaction policies are injected into the model. What unifies these methods is not a common architecture, but a common aim: to make the computational path to diagnosis more clinically aligned, more inspectable, and more robust than generic prompting allows.