DiagnosticSLM: Multi-Domain Diagnostic Systems

• DiagnosticSLM is a label for a family of systems that convert raw observations into structured representations for tasks such as classification, decision support, latent-state recovery, or model auditing.
• It spans diverse applications including label-free breast tissue screening with SLIM, sequential clinical diagnosis using language models, automotive fault detection via specialized transformers, latent attribute analysis, and PDE simulator auditing.
• These systems share a design pattern that emphasizes actionable intermediate diagnostics, enabling improved accuracy and consistency across varied technical domains.

Searching arXiv for papers using the term "DiagnosticSLM" and closely related titles to ground the article. DiagnosticSLM is not a single canonical method but a label that has been attached to several technically unrelated systems. In the cited literature, it refers to: a label-free breast-tissue screening pipeline built on Spatial Light Interference Microscopy (SLIM); a sequential clinical diagnosis framework for LLMs centered on the Sequential Diagnosis Benchmark and the MAI Diagnostic Orchestrator; a 3B-parameter domain-specific small LLM for automotive fault diagnosis; a joint maximum-likelihood framework for large-scale structured latent attribute analysis; and a diagnostic software suite for auditing learned PDE simulators as approximate evolution operators (Majeed et al., 2017, Nori et al., 27 Jun 2025, Kumar et al., 23 Nov 2025, Gu et al., 2020, Shikhman, 16 Jun 2026). The shared term therefore denotes a family of domain-specific diagnostic systems rather than a unified research program.

1. Terminological scope and recurrent design pattern

The term has been used in at least five distinct technical settings. The commonality is functional rather than architectural: each system converts raw observations into a structured diagnostic representation, then uses that representation for classification, decision support, latent-state recovery, or model auditing.

Usage of “DiagnosticSLM”	Domain	Core mechanism
SLIM-based screening	Breast histopathology	Optical path-length biomarkers + LDA
Sequential diagnosis with LMs	Clinical reasoning	SDBench + MAI-DxO
Domain-specific small LM	Automotive diagnostics	Guided data generation + DAPT → DSFT → DPO
Joint-MLE latent attribute analysis	Educational/psychometric inference	Joint estimation of $Q$, $A$, and $\Theta$
PDE simulator auditing suite	Scientific machine learning	Structural diagnostic panel beyond relative $L^2$

A recurrent pattern is visible across these uses. Each version defines a compact diagnostic state: a phase-derived feature vector in pathology, a probability-ranked differential in sequential medicine, an in-domain instruction-following model state in industrial diagnosis, a binary attribute profile in latent-variable analysis, or a panel of structural consistency metrics in learned simulation. This suggests that the label has functioned as a domain-local shorthand for systems that emphasize diagnostically actionable intermediate structure rather than end-to-end black-box scoring.

2. SLIM-based breast-tissue screening

In the breast-pathology setting, DiagnosticSLM denotes a quantitative method for label-free tissue screening using Spatial Light Interference Microscopy. A SLIM module (CellVista SLIM Pro, Phi Optics) is coupled to the output port of a commercial phase-contrast microscope (Zeiss Axio Observer Z1); a $40\times/0.75$ NA phase-contrast objective collects light; the image plane is relayed onto a CCD (Andor Zyla) via a $4f$ system with lenses $L_1$ and $L_2$; and, at the Fourier plane between $L_1$ and $L_2$, a phase-only spatial light modulator imposes four phase shifts $A$0 between the unscattered and scattered fields. From the four intensity frames $A$1, $A$2, $A$3, and $A$4, the phase map is reconstructed by phase stepping. The measured phase satisfies

$A$5

with $A$6, $A$7, and $A$8–$A$9. Optical-path length is then

$\Theta$0

after subtraction of a background phase map and optional smoothing (Majeed et al., 2017).

The diagnostic representation is gland-centric. Three classes of biomarkers are extracted from the phase map $\Theta$1. The geometric feature is gland perimeter curvature, computed on epithelial regions manually outlined in ImageJ on co-registered H&E images. For a perimeter parameterization $\Theta$2, the extrinsic curvature is

$\Theta$3

and the gland-level statistic is the median curvature $\Theta$4. The scattering feature is mean scattering length, based on the scattering-phase theorem,

$\Theta$5

with a $\Theta$6 pixel window and gland-level median $\Theta$7. Texture is captured by Leung-Malik filter-bank responses $\Theta$8, $\Theta$9, clustered by K-means with $L^2$0 to define textons; local $L^2$1 neighborhoods yield normalized histograms, and the gland-level feature is the median of each texton bin, forming a 50-dimensional vector $L^2$2. Each gland is therefore represented by the 52-dimensional vector

$L^2$3

Classification is performed with Linear Discriminant Analysis with equal prior class probabilities and default MATLAB LDA, without additional regularization. The study used a tissue microarray (US Biomax BR-1002), $L^2$4 FFPE sections, and $L^2$5 cores. Sixty-eight cores were selected: 34 malignant (IDC) and 34 benign (normal or tumor-adjacent normal), each containing intact epithelium. Three-fold cross-validation used non-overlapping folds of approximately 23 cores each. Gland-wise likelihood scores were averaged per core to obtain a core score $L^2$6, and ROC analysis yielded an AUC of 0.91. At the optimal operating point with equal cost assigned to false positives and false negatives, sensitivity was 0.94, specificity 0.85, and accuracy approximately 0.895. The system is described as instrument-independent because it measures intrinsic optical path-length rather than stain-dependent contrast, and the combination of full-slide scanning, quantitative phase biomarkers, and machine learning is positioned as a decision-support mechanism to reduce observer variability and accelerate histopathology workflows.

3. Sequential clinical diagnosis with LLMs

In the clinical-language-model setting, DiagnosticSLM denotes a cost-aware sequential diagnostic framework rather than a static question-answering benchmark. The Sequential Diagnosis Benchmark transforms 304 diagnostically challenging NEJM Clinicopathological Conference cases, published between 2017 and 2025, into multi-turn encounters. Each case begins with a 2–3 sentence chief-complaint vignette. At each turn, the diagnostic agent may ask one or more free-text questions about history or physical examination, order one or more explicit diagnostic tests, or commit to a final diagnosis. A gatekeeper model, implemented as o4-mini with access to the full case file, reveals only findings legitimately obtainable from the requested query or test; for unpublished details it synthesizes plausible, numerically consistent findings; and it refuses to provide interpretation, diagnostic impressions, or pathognomonic hints ahead of confirmatory testing. Physician review of 508 sampled responses found that only 8 raised any flag and none leaked the diagnosis. Final diagnoses are graded by a Judge LM on a 1–5 Likert scale, with scores $L^2$7 counted as correct; Cohen’s $L^2$8 against human adjudicators is 0.70–0.87. Cost per case is defined as

$L^2$9

with \$300 per physician visit plus CPT-based test costs (Nori et al., 27 Jun 2025).

The operational core is the MAI Diagnostic Orchestrator (MAI-DxO), a model-agnostic orchestration layer that simulates a five-person virtual panel using specialized prompts over a single underlying LM. Dr. Hypothesis maintains a probability-ranked top-3 differential and updates $40\times/0.75$0 in a Bayesian manner after each result. Dr. Test-Chooser selects up to three tests or questions that maximally discriminate among leading hypotheses. Dr. Challenger searches for anchoring and suggests disconfirming tests. Dr. Stewardship vetoes expensive low-yield actions and proposes cheaper alternatives. Dr. Checklist validates test names and internal consistency. The loop alternates Bayesian differential maintenance,

$40\times/0.75$1

diagnosis commitment when $40\times/0.75$2, and value-of-information test selection. Utility is written as

$40\times/0.75$3

with entropy $40\times/0.75$4, and the test chooser computes

$40\times/0.75$5

ranking candidate tests by $40\times/0.75$6. Heuristics cap the process at up to 5 questions per turn and up to 3 tests per turn, and a budgeted variant cancels tests that exceed the remaining budget.

Performance is reported along an accuracy–cost frontier. Twenty-one human physicians with median 12 years of experience, evaluated on 56 held-out cases, achieved 19.9% accuracy at an average cost of \$40\times/0.75$74k–\$40\times/0.75$$87,850 per case. MAI-DxO with o3 achieved 81.9$$40\times/0.75$92,396 in cost-focused mode, and 85.5$4f$0p<0.005$. The stated limitations are also substantial: the NEJM CPC corpus is skewed toward rare and complex conditions, the cost model approximates U.S. prices while omitting invasiveness and logistics, there is no explicit model of false positives or false negatives in healthy or benign cases, and the physician baseline excludes external resources.

4. DiagnosticSLM as a domain-specific small LLM for automotive diagnostics

In the industrial setting, DiagnosticSLM denotes a 3B-parameter transformer specialized for fault diagnosis, root-cause analysis, and repair recommendation in the automotive domain. Its foundation is Llama-3.2-3B, with 32 transformer layers, hidden dimension 4096, MLP dimension 11008, 32 attention heads, rotary embedding dimension 128, and a 32k-token vocabulary. Total parameter count is 3,000,000,000. FP16 weights occupy approximately 6 GB; with AdamW states and optimizer buffers, peak GPU memory is approximately 18 GB; and after full-shard quantization to 4 bits, the model can be served in approximately 4 GB. Training uses PyTorch Fully Sharded Data Parallel across two NVIDIA RTX 4090 GPUs, and the final preference-optimization stage uses LoRA with rank 16 and $4f$1, injecting approximately 1 million trainable parameters (Kumar et al., 23 Nov 2025).

The data pipeline is a central part of the system. Keyphrases are extracted from internal technical manuals using a 4-bit quantized Llama-3-70B-Instruct. Automated Google Custom Search over approximately 200 automotive keyphrases yields 706,971 webpages and 403M tokens. A random sample of 20,000 pages is labeled by Llama-3-70B-Instruct as relevant or irrelevant, producing 11,621 relevant and 8,379 irrelevant labels in 22 GPU-hours; logistic regression with L2 regularization, $4f$2, and solver SAGA then reaches 88.2% test accuracy on an 80/20 split and classifies the full set, with 356,312 pages deemed automotive-related. Bottom-up domain curation creates eight topic documents, computes cosine similarity between each page and each topic embedding, retains pages with

$4f$3

recovers false negatives from the non-relevant pool using the same threshold and top-20% retention, and after MinHash deduplication produces 387,572 unique samples comprising approximately 257M tokens. Guided synthetic augmentation uses Gemma-2-27B as teacher, with a prompt that removes non-automotive sentences and expands relevant passages with factually accurate details and comprehensive explanations; this stage consumes approximately 5400 GPU-hours and yields a final 206M-token automotive corpus after post-augmentation deduplication. During DSFT data generation, topic and task sampling are uniform:

$4f$4

Training proceeds in three stages: DAPT, DSFT, and DPO. DAPT uses causal language modeling,

$4f$5

on the 206M-token automotive corpus with AdamW $4f$6, cosine decay with 10% linear warmup, peak learning rate $4f$7, per-device batch size 1, gradient accumulation 8, global batch size 16, sequence length 2048, FP16 precision, 5789 steps, and approximately 118 GPU-hours. DSFT uses

$4f$8

over DiagnosticMix, a 72,000-pair automotive-and-Alpaca instruction dataset, with the same optimizer and schedule but learning rate $4f$9, per-device batch 2, gradient accumulation 8, global batch 32, 2130 steps, and approximately 47 GPU-hours. DPO uses

$L_1$0

with $L_1$1, UltraFeedback-Binarized comprising approximately 64,000 preference pairs, LoRA fine-tuning, AdamW, learning rate $L_1$2, cosine decay with 10% warmup, per-device batch 1, sequence length 2048, bfloat16, one epoch, and approximately 144 GPU-hours.

Evaluation is organized into four test-only benchmarks. DiagnosticMCQ contains 876 four-option multiple-choice questions evaluated with 5-shot prompting and standard accuracy. DiagnosticQA converts these questions to free-form QA. DiagnosticComp contains 117 sentence-completion items scored by total log-likelihood,

$L_1$3

choosing the highest-scoring option. DiagnosticSum contains 200+ technical explanations paired with two-line summaries and reports ROUGE-1/2/L, BLEU, BERTScore-F1, and cosine similarity. On DiagnosticMCQ, DiagnosticSLM scores 45.32%, versus 36.53% for Llama-3.2-3B-Instr., 40.98% for Phi-4-mini-instr., 37.33% for Gemma-2-2B-instr., 46.92% for Llama-3.1-8B-instr., and 45.21% for Gemma-2-9B-instr. The reported gain over the closest 3B baseline is +8.79 percentage points, approximately 25% relative. On DiagnosticQA it reaches 38.31%, and on DiagnosticComp 55.56%. In summarization, however, Phi-4-mini-instr. exceeds it across ROUGE, BLEU, BERTScore, and cosine similarity. The paper therefore presents DiagnosticSLM as a strong in-domain reasoner and generalizer within automotive diagnostics, while explicitly noting weaker summarization performance, possible factual drift from synthetic data, and uncertainty outside the eight ASE topics. Additional stated properties include low latency, reported as under 50 ms/token on GPU, and on-premise inference for data privacy.

5. Joint-MLE structured latent attribute analysis

In the latent-variable setting, DiagnosticSLM refers to a framework for structured latent attribute models, where the observed data are an $L_1$4 binary matrix $L_1$5, each subject $L_1$6 has a binary latent attribute vector

$L_1$7

and a $L_1$8 binary $L_1$9-matrix specifies which attributes each item depends on. Conditioned on $L_2$0, $L_2$1, and item parameters $L_2$2, responses are Bernoulli with success probability $L_2$3. In the two-parameter DINA special case,

$L_2$4

with $L_2$5. In multi-parameter variants such as GDINA or LCDM, the logit can include main effects and higher-order interactions over subsets of the active attributes (Gu et al., 2020).

The central departure from standard marginal-likelihood treatments is to regard both the latent attribute profiles $L_2$6 and the $L_2$7-matrix as fixed unknowns and to estimate them jointly with the continuous item parameters. The generic Bernoulli log-likelihood is

$L_2$8

and the joint MLE is

$L_2$9

subject to model-specific constraints such as $L_1$0 in DINA or sparsity restrictions on $L_1$1. For fixed $L_1$2 in the two-parameter case, the item-parameter MLEs reduce to empirical proportions over the ideal-response clusters:

$L_1$3

The asymptotic theory is formulated under triple asymptotics $L_1$4. The assumptions include response probabilities bounded away from 0 and 1, a positive separation gap between positive and negative response probabilities, many copies in $L_1$5 of each standard-basis vector, sufficient frequency of each of the $L_1$6 attribute profiles, and moderate growth conditions such as $L_1$7 and $L_1$8. Defining

$L_1$9

Theorem 1 shows vanishing average squared error in estimated response probabilities and, up to a column permutation of $L_2$0, vanishing item-wise and subject-wise classification errors for $L_2$1 and $L_2$2. In the regime with $L_2$3 and fixed $L_2$4, the rate becomes $L_2$5.

Computation is organized around ADG-EM for the two-parameter case and a two-stage extension for multi-parameter models. ADG-EM alternates an approximate E-step, which draws $L_2$6 Gibbs samples for each binary latent variable and updates running averages via SAEM weighting, with an M-step that thresholds the running averages to obtain $L_2$7 and $L_2$8, constructs the ideal-response indicator $L_2$9, and updates $A$00 and $A$01. The per-iteration computational cost is $A$02, effectively $A$03 when $A$04. For multi-parameter SLAMs, stage 1 runs ADG-EM under a DINA approximation to recover $A$05 and $A$06, and stage 2 performs regression refinement using screened interaction features and penalized logistic regression. Simulations with $A$07 up to 2000 and $A$08 up to 15 are reported to recover $A$09 perfectly or within a few errors in approximately 10 iterations; for $A$10, exact recovery occurred in 100% of 200 replications. On TIMSS 2011 Austrian assessment data with $A$11, $A$12, $A$13, and approximately 52% missingness by design, the estimated $A$14 matched 37 of 47 rows exactly, while additional dependencies for the remaining items were described as aligning with pedagogical intuition.

6. DiagnosticSLM as a software suite for learned PDE simulators

In scientific machine learning, DiagnosticSLM is a post hoc auditing suite for learned PDE simulators viewed as approximate evolution operators rather than mere state predictors. Its software contract requires exactly four inputs: reference trajectories on a fixed spatial grid; either a learned propagator $A$15 or saved predictions; equation metadata specifying available physical structures such as an energy functional $A$16, a flux $A$17, a source $A$18, an admissible manifold $A$19, and scaling transforms; and a diagnostic configuration specifying the diagnostics, horizons, control volumes, and perturbation protocols to run. These inputs are normalized into a TrajectoryDataset and processed by DiagnosticRunner, which queries a ModelWrapper, checks applicability through metadata, and aggregates results into a DiagnosticReport (Shikhman, 16 Jun 2026).

The package emphasizes that relative state error is necessary but insufficient. Its baseline metric is relative $A$20 error,

$A$21

Structural coherence is then audited through semigroup consistency,

$A$22

finite-difference generator discrepancy,

$A$23

with normalized comparison across short horizons, energy-balance diagnostics for conservative and dissipative systems, integral conservation or balance over control volumes,

$A$24

admissibility or constraint violation,

$A$25

perturbation-response mismatch between model and reference trajectories, and scaling-law consistency under known similarity transforms. The point of this panel is to test temporal composition, local generator fidelity, physical conservation laws, admissibility, and sensitivity structure.

The software architecture is built from four core classes and a registry of diagnostics. TrajectoryAdapter loads HDF5, NumPy, or NetCDF trajectories, restores grid and normalization metadata, and emits a TrajectoryDataset. ModelWrapper provides a unified predict(u0,k) interface for both one-step propagators and precomputed prediction arrays. EquationMeta stores callables such as energy(u), flux(u), source(u), constraint projectors, and scaling transforms. DiagnosticRunner traverses horizons and trajectories, calls the registered kernels, and collects the raw metric arrays. This design makes equation-specific diagnostics conditional on supplied metadata while keeping relative error, semigroup tests, generator discrepancy, and perturbation response broadly applicable.

Validation covers five benchmark PDE tasks—2D incompressible Navier–Stokes, 2D shallow-water dynamics, 2D active matter, 3D compressible Navier–Stokes, and 3D magnetohydrodynamics—and four surrogate architectures: FNO, DeepONet, U-Net, and a ResNet-style CNN. Each architecture is studied in well-trained, underfit, and oversmoothed variants. The central empirical claim is that structural diagnostics can deteriorate substantially while relative $A$26 error remains moderate or even improves. In SW2D with an FNO surrogate, the well-trained model has $A$27, $A$28, and $A$29; the underfit variant has $A$30, $A$31, and $A$32; and the oversmoothed variant has $A$33, $A$34, and $A$35. In MHD3D oversmoothed ResNet, the energy-decay error triples while relative error increases by only approximately 40%. In underfit U-Net on AM2D, relative error improves from 0.747 to 0.682 even as semigroup error rises from 0.201 to 0.317, a reported increase of 58%. The suite is therefore intended for regression testing, model selection, and deployment auditing in settings where structural fidelity matters at least as much as pointwise prediction error.