Neurosymbolic Alignment for Physiologically-Safe Clinical Language Models
Authors: Abdulhady Abas Abdullah, Erik Cambria, Milena Zivkovic
Organizations: Artificial Intelligence and Innovation Centre, University of Kurdistan Hewlêr, Erbil, Iraq · MIT Media Lab, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Department of Physics, Faculty of Science, University of Kragujevac, 34000 Kragujevac, Serbia
Clinical LLMs can generate recommendations that are factually plausible yet physiologically unsafe. We investigate whether safety alignment can be improved by grounding preference optimization in structured physiological knowledge rather than text-only supervision. Methods: We propose Neurosymbolic Alignment, a training-time framework that couples a 7B clinical LLM with an HGNN-based Physiological World Model over an 847K-node biomedical knowledge graph. Candidate responses are scored using homeostatic constraints, multi-hop path plausibility, and drug-interaction penalties, and the resulting rankings drive iterative on-policy ORPO updates. Evaluation is performed on the Clinical Safety Benchmark (CSB), a 2,500-scenario benchmark for physiological constraint violations in generative clinical reasoning. Results: Relative to ORPO, the proposed method improves CSS from 69.5% to 90.8% (+21.3 pp), reduces physician-evaluated HR from 14.1% to 5.1% on the blinded subset, and improves DID from 72.8% to 91.6%. These gains are corroborated by an HGNN-independent Rule-Engine Safety Score (RSS: 86.4%, +21.2 pp over ORPO; r=0.97 concordance with CSS). The method also exceeds GPT-4 (5-shot) on all safety metrics despite a 10x parameter disadvantage, and outperforms an inference-time self-correction pipeline (SFT+SelfCorrect) by 11.4 pp CSS. Under synthetic EHR-style noise, 84.2% CSS is retained. Ablation analysis shows that HGNN scoring (-16.2 pp) and iterative training (-11.5 pp) are the dominant contributors. PhysioScore calibration against 200 clinician labels yielded ECE = 0.038 and kappa = 0.91. Conclusion: Training-time physiological grounding produces measurable and independently verifiable safety improvements in open-weight clinical LLMs under controlled evaluation. External validation on real clinical data is required to determine whether these gains transfer to deployment settings
Figures & tables
Fig. 1: Training-time neurosymbolic pipeline. The base LLM samples candidate clinical responses, the Physiological World Model scores their feasibility, and Iterative ORPO updates the policy. At inference, only the aligned LLM is used.
Fig. 2: Training-time neuro-symbolic alignment: the LLM samples candidates, the HGNN scores physiological feasibility on G , and Iterative ORPO updates the policy (HGNN not used at inference).
Source
Entities
Relations
Triples
UMLS Metathesaurus
4,200,000
54
12,847,392
DrugBank 5.1
14,315
12
1,432,847
Reactome
12,632
8
284,521
HPO
16,234
5
156,842
SIDER
1,430
3
139,756
PharmGKB
5,421
7
87,432
TABLE I: Graph Source Statistics
Node Type
Count
Percentage
Drug
14,315
1.69%
Disease
24,166
2.85%
Symptom
7,823
0.92%
Organ/Anatomy
12,478
1.47%
Pathway
12,632
1.49%
Gene/Protein
19,566
2.31%
TABLE II: Node Type Distribution in Physiological World Model
Method
Base Model
Training Data
KG
Pref. Opt.
SFT
Mistral-7B
MedQA + PubMed
✗
✗
SFT+RAG (lightweight)
Mistral-7B
MedQA + PubMed
Retrieval
✗
SFT+DenseRAG
Mistral-7B
MedQA + PubMed
Dense+Rerank
✗
DPO
Mistral-7B
Human Prefs (5K)
✗
✓
ORPO
Mistral-7B
Human Prefs (5K)
✗
✓
KG-LLM
Mistral-7B
MedQA + KG
Static
✗
TABLE III: Baseline Method Specifications
Method
MA
CSS
RSS
HR ↓
DID
PC
SFT
61.2
58.4
54.6
23.7
62.1
0.61
SFT+RAG
67.8
63.2
60.8
68.5
0.67
SFT+DenseRAG
70.8
78.3
74.2
81.4
0.78
DPO
69.4
67.8
63.4
71.3
0.69
ORPO
70.1
69.5
65.2
14.1
72.8
0.71
KG-LLM
68.9
72.4
68.1
12.8
76.2
0.74
TABLE IV: Performance comparison across methods. HR is reported only for the blinded physician-evaluated subset. RSS is HGNN-independent (DrugBank-Rule deterministic checker).
Claim
Result
Ref.
Indep.
CSS over ORPO
+21.3 pp
Tab. IV
H
RSS over ORPO
+21.2 pp
Tab. IV
I
DID over ORPO
+18.8 pp
Tab. IV
I
HR over ORPO (physician)
− 9.0 pp
Tab. IV
P
CSS vs. GPT-4 (5-shot)
+10.2 pp
Tab. IV
H
CSS vs. SFT+SelfCorrect
+11.4 pp
Tab. IV
H
TABLE V: Summary of Key Demonstrated Results. Metric independence: I = fully HGNN-independent; H = HGNN-derived; P = physician-adjudicated.
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Fig. 3: Schema of the heterogeneous physiological knowledge graph showing the six node types ( TV ) and five canonical relation types ( TE ).
Fig. 4: Performance progression across training iterations. CSS and DID show the steepest improvements in early iterations, while gains diminish after iteration 4. The inverted hallucination rate (100-HR) demonstrates consistent safety improvements throughout training.
TABLE XI: PhysioScore Combination Function Sensitivity
λ Perturbation
CSS
Δ CSS
HR ↓
Δ HR
Note
Default λ values (Table 5)
90.8
5.1
All λ×0.5 (halved steepness)
87.3
− 3.5
7.4
+ 2.3
Softer penalties
All λ×2.0 (doubled steepness)
89.2
− 1.6
5.9
+ 0.8
Harsher but noisier
Uniform λ=2.0 for all constraints
88.1
− 2.7
6.8
+ 1.7
Ignores clinical risk profiles
λ from Gaussian noise ( ±20% )
89.9
− 0.9
5.5
+ 0.4
Robust to moderate noise
Appendix
TABLE XII: Steepness Parameter ( λ ) Sensitivity on CSS
Iteration
MA (%)
CSS (%)
HR (%) ↓
DID (%)
PC
Iteration 1
71.3
78.2
10.4
81.5
0.79
Iteration 2
72.4
82.6
8.9
84.2
0.82
Iteration 3
73.8
86.4
7.2
87.3
0.85
Iteration 4
74.6
89.1
5.8
90.1
0.87
Iteration 5
75.2
90.8
5.1
91.6
0.89
Δ (Iter 1 → 5)
+3.9
+12.6
-5.3
+10.1
+0.10
Appendix
TABLE XIII: Performance Progression Across Training Iterations
Specialty
CSS
CSS
Δ
HR
DID
(Ours)
(ORPO)
(Ours)
(Ours)
Cardiology
91.2
70.3
+20.9
4.8
92.4
Nephrology
89.7
68.1
+21.6
5.3
90.8
Oncology
87.4
65.8
+21.6
6.1
88.2
Pediatrics
86.9
64.2
+22.7
5.8
89.1
Psychiatry
84.3
67.5
+16.8
7.2
82.6
Appendix
TABLE XIV: Cross-Domain Generalization Results
Fig. 5: Training loss convergence over 30K steps. Total loss decreases from 2.3 to 0.38, with KG loss showing the fastest convergence. The OR loss initially increases as preference learning intensifies, then stabilizes.
Metric
Iter 1
Iter 2
Iter 3
Iter 4
Iter 5
Total Loss
0.82
0.61
0.48
0.41
0.38
SFT Loss
0.95
0.78
0.68
0.62
0.60
KG Loss
0.28
0.18
0.14
0.12
0.11
OR Loss
0.35
0.32
0.30
0.29
0.28
Pref. Pairs Generated
1,750
1,750
1,750
1,750
1,750
Avg. Score Gap
0.42
0.38
0.31
0.26
0.22
Appendix
TABLE XV: Training Statistics Across Iterations
Component
Parameter
Value
LLM Training
Base Model
Mistral-7B-v0.3
Learning Rate
2×10−5
LR Schedule
Cosine Decay
Batch Size
32
Max Sequence Length
2048
Gradient Accumulation
4
Appendix
TABLE XVI: Hyperparameter Configuration
Stage
GPU Hours
Memory (GB)
Storage (GB)
KG Construction
2.4
128
45
HGNN Pre-training
8.6
320
12
LLM Fine-tuning (per iter.)
3.2
640
28
Response Generation
1.8
160
8
Preference Scoring
0.4
64
2
Total (5 iterations)
42.6
95
Appendix
TABLE XVII: Computational Resource Requirements
Dataset
Train
Valid
Test
Task Type
MedQA-USMLE
10,178
1,272
1,273
Multiple Choice
PubMedQA
450
50
500
Yes/No/Maybe
MedMCQA
182,822
4,183
6,150
Multiple Choice
MMLU-Medical
1,089
123
1,871
Multiple Choice
CSB (Ours)
1,750
250
500
Open Generation
DDI-Corpus
4,521
645
1,290
Interaction Detection
Appendix
TABLE XVIII: Dataset Statistics and Characteristics
Stronger KL regularization and score-gap calibration
Linguistic Hallucinations
56
2.8
Temporal fabrication outside graph representation
Temporal constraints and post-hoc consistency checks
Dosing Errors
41
2.1
Incomplete pharmacokinetic data
Weight/renal-aware dosing calculators in tool loop
Context Ambiguity
38
1.9
Insufficient patient information
Uncertainty-aware abstention and clarification prompts
Novel Drug Combinations
29
1.5
Emerging pharmacological knowledge
Rapid ingestion of approvals and interaction bulletins
Appendix
TABLE XXI: Taxonomy of Failure and Root Cause Analysis (training+validation scenarios)
Fig. 6: Step-wise toy example of constraint verification. A candidate plan is generated, entities are linked to the physiological graph, and HGNN-based checks reject the plan due to hyperkalemia-related contraindications.
Large language models (LLMs) are increasingly used for medical and health-related questions, yet their safety in high-risk medical scenarios remains poorly understood. We introduce MedHarm\footnote{Code and data will be released upon acceptance. Due to the sensitive nature of high-risk medical queries, data access will be available to qualified researchers upon request.}, a high-risk medical safety benchmark with 1,100 medically grounded queries across 10 safety-critical categories, including toxicology, pharmacology, covert poisoning, anesthesia, and fetal harm. Unlike broad medical QA benchmarks, MedHarm targets realistic clinical, educational, and technical prompts that require refusal, caution, or safe redirection rather than direct helpfulness. We evaluate 15 LLMs spanning general-purpose, medical-purpose, closed-source, and downstream SFT models, together with 4 representative guardrail models. Results reveal a substantial gap between apparent alignment and medical safety: aligned models can still produce unsafe or actionable responses, medical fine-tuning can amplify harmful specificity, and external guardrails reduce some failures while introducing brittle blocking and weak safe helpfulness. These findings show that medical safety cannot be inferred from general alignment or medical capability alone, highlighting the need for domain-specific stress testing before deploying LLMs in safety-critical medical applications.
Yige Li, Jun Sun, Wei Zhao +5
Singapore Management University · Deakin University · The University of Melbourne +2
Large language models(LLMs) increasingly match expert performance on licensing examinations, yet routine clinical use remains limited because governance requires auditable reasoning, safety and ethics alignment, and resilience to adversarial misuse. Here we present SafeMed-R1, trained with a traceable Clinical Trust Signals(CTS) pipeline that links each reasoning instance to clinician rubric scores and edit histories, and aligned through safety and ethics supervision and red team stress testing. SafeMed-R1 attains a macro-averaged accuracy of 79.6% across clinical benchmarks. Under adversarial safety testing, it shows the lowest aggregated risk and reduces unsafe outputs by about 3 to 5% relative to its baseline. In a paired expert study of 30 medication safety vignettes, SafeMed-R1 matches PGY1 and PGY2 residents on medical correctness and scores higher for medication safety, guideline consistency, and clinical usefulness. Collectively, these results suggest that clinician-audited supervision provenance, together with domain-tailored safety and ethics alignment, can strengthen governance-relevant evidence without relying on inference-time retrieval or citation grounding.
Chao Ding, Mouxiao Bian, Tianbin Li +12
Shanghai Artificial Intelligence Laboratory, Shanghai, China · Joint Laboratory of Biomedical Artificial Intelligence, Shanghai East Hospital, Tongji University School of Medicine, Shanghai, China · School of Computer Science and Technology, Tongji University, Shanghai, China +5
Large language models (LLMs) have become significant providers of mental health support, yet they remain products of an attention economy whose operational and commercial targets favour sustained engagement over the friction that effective psychological support often requires. Developers' safety responses have been largely reactive, addressing the most visible and acute harms while subtler, longer-term patterns of risk (e.g., dependency, boundary erosion, the amplification of distorted beliefs) receive less attention. We contend that making LLMs structurally safe requires alignment organised at three levels that mirror how society assures the safety of human clinical practice: 1) explicit value specification grounded in the codified normative commitments of clinical practice; 2) training that embeds those values in the model; and 3) oversight that detects drift and longer-term harm during deployment, much as clinical supervision does for human practice. Organising alignment in this way yields a construct we call alignment plausibility - a structured demonstration that a system's values, training regime, and oversight mechanisms are together consistent with safe and positive outcomes. We propose alignment plausibility as a regulatory construct (by drawing analogy to the established construct of biological plausibility) for AI in health: a principled way to argue for, or against, trust that systems are aligned to positive health outcomes, will cause no harm even where capable of doing so, and will ultimately lead to patient benefit.