Alpha-TCM: A Nine-Layer Framework for TCM-Guided Multimodal Inference and Structured-Prior Auditing

Yucheng Liu; Everett X. Wang

cs.LGOct 28, 2026

Alpha-TCM: A Nine-Layer Framework for TCM-Guided Multimodal Inference and Structured-Prior Auditing

Authors: Yucheng Liu, Everett X. Wang

Organizations: Department of Engineering Science, University of Oxford, Oxford, United Kingdom · Beijing Institute of Nanoenergy and Nanosystems, Chinese Academy of Sciences, Beijing, China

Abstract

Alpha-TCM makes two contributions under a central thesis: auditable systems should separate domain knowledge, case evidence, probabilistic inference, and downstream authority. First, it proposes a nine-layer TCM-oriented architecture covering semantic definitions, multimodal acquisition, quality control, provenance-bearing knowledge, probabilistic inference, longitudinal updating, constrained decisions, language interaction, and human governance. Each layer specifies its information contract, failure boundary, dependencies, and validation status. This is a system-design proposal, not a claim of complete clinical implementation. Second, a controlled synthetic prototype instantiates Layers 1-5 to audit the knowledge-evidence interface. Four conditional-Gaussian encoders process synthetic tongue images, pulse waveforms, inquiry fields, and text, while a frozen bridge predicts a five-dimensional synthetic outcome. A versioned contextual prior is compiled independently of the renderer. Fusion removes duplicated reference-prior information before adding context. Equal-information controls, prior regularization, graded contamination, and structural corruption test contextual benefits, Bayesian specificity, and conflict detection. Across ten paired seeds and four training sizes, contextual knowledge improved mean absolute error over marginal committee fusion at 600 and 1,200 cases, but not over equal-information neural controls. Fully incorrect context harmed performance at every size, while prior effects were inconsistent and conflict scores did not track case-level damage. The results support contextual structure in low-data synthetic settings, not a specifically Bayesian advantage. The prototype establishes no claims about TCM theory or clinical, diagnostic, therapeutic, causal, longitudinal, or transfer validity.

Figures & tables

Req.	System invariant	Failure prevented	Verification obligation
R1	Every construct has an operational definition, source, version, and epistemic status	Terminology treated as biological fact	Semantic and provenance audit
R2	Observation, availability, validity, quality, and negative evidence remain distinct	Missing or invalid input treated as patient state	Acquisition and missingness tests
R3	Knowledge and case evidence use separate interfaces	Domain theory copied into observations or labels	Source and data-flow traceability
R4	Posterior state, downstream prediction, decision, and explanation remain distinct	Prediction silently converted into diagnosis or action	Interface and authority tests
R5	Prior influence is measurable, challengeable, and removable	Structured knowledge becomes unfalsifiable authority	Absent, aligned, and wrong-prior controls
R6	Temporal conditioning and causal claims require separate assumptions	Association presented as treatment effect	Longitudinal and causal-design audit

Table 1: System requirements that remain binding across all nine Alpha-TCM layers. They are architectural invariants, not claims of completed clinical validation.

Panel of Figure 1 — Figure 1: Alpha-TCM’s nine-layer dependency order. The architecture covers the complete information pathway; the evaluated prototype instantiates only the static Layers 1–5 vertical slice.

Layer	Principal contract	Failure contained	Evaluated status
1	Versioned identifiers, definitions, sources, and epistemic status	Semantic ambiguity and category collapse	Synthetic contract implemented
2	Raw observations plus acquisition metadata	Modality and acquisition conflation	Four synthetic modalities
3	Separate availability, validity, quality, and negative findings	Missingness treated as evidence	Inquiry mask only
4	Versioned, challengeable relations and prior potentials	Theory treated as unchangeable fact	Contextual prior compiled
5	Conditional evidence, prior accounting, posterior state, uncertainty	Overconfident or duplicated fusion	Encoders, BCM, contextual fusion
6	Transition, filtering, smoothing, and forecasting contracts	Visits treated as independent snapshots	Specified, not evaluated

Table 2: Nine-layer contracts, the failures they contain, and the exact boundary of the current prototype.

Layer	Prototype artifact	Tested property
1	Versioned synthetic taxonomy, latent-mapping hypotheses, and source registry	Definitions and provenance are explicit and hashable
2	Tongue-like image, pulse-like waveform, masked inquiry, text, and context metadata	Heterogeneous observations share a case without sharing an encoder
3	Inquiry observation mask and modality omission interface	Missing information is represented separately from observed values
4	Contextual-prior contract and deterministic knowledge compiler	Knowledge is separate from renderer code and can be corrupted or removed
5	Four conditional experts, marginal and contextual fusion, bridge, uncertainty, and conflict audit	Prior influence, prediction, uncertainty, and disagreement are separately measurable

Table 3: Traceability from the nine-layer architecture to the implemented static prototype.

Modality	Input	Encoder	Evidence output
Tongue	RGB image	Three strided 2-D convolutions and global pooling	$4$ means, $4$ variances
Pulse	Length-256 waveform	Residual 1-D CNN and global pooling	$4$ means, $4$ variances
Inquiry	20 codes plus mask	Two-layer masked MLP, hidden width 64	$4$ means, $4$ variances
Text	Up to 384 UTF-8 bytes	32-D byte embedding; kernels 3, 5, 7; max pooling	$4$ means, $4$ variances

Table 4: Implemented modality encoders. Descriptions follow the executable model definitions; they are deliberately compact and CPU-sized.

Train $n$	Marginal BCM	Knowledge context	Learned context	Late context
600	$0.2340\pm 0.0057$	$0.2329\pm 0.0052$	$0.2329\pm 0.0045$	$0.2340\pm 0.0046$
1,200	$0.2205\pm 0.0034$	$0.2200\pm 0.0033$	$0.2193\pm 0.0037$	$0.2174\pm 0.0036$
2,400	$0.2071\pm 0.0043$	$0.2069\pm 0.0044$	$0.2048\pm 0.0036$	$0.2026\pm 0.0036$
4,800	$0.1955\pm 0.0036$	$0.1952\pm 0.0035$	$0.1937\pm 0.0034$	$0.1931\pm 0.0038$

Table 5: Confirmation MAE for the four shared-encoder systems. Lower is better. Context features are matched across the knowledge, learned-context, and late-context systems.

Figure 2: Confirmation learning curves across ten paired seeds. Absolute MAE decreases with training size. The researcher-specified synthetic contextual prior has a small low-data advantage over marginal BCM, whereas the equal-information learned and late-context systems become stronger as sample size increases.

Figure 3: Prespecified paired MAE contrasts with bootstrap 95% intervals. Negative values favor knowledge context for the first three contrasts; positive values in the wrong-minus-aligned contrast indicate harm. An interval crossing zero is reported as inconclusive, not equivalent.

Train $n$	MAE	RMSE	Gaussian NLL	80% cov.	95% cov.
600	$0.2329\pm 0.0052$	$0.2942\pm 0.0047$	$0.2049\pm 0.0283$	$0.760\pm 0.018$	$0.922\pm 0.012$
1,200	$0.2200\pm 0.0033$	$0.2798\pm 0.0030$	$0.1167\pm 0.0232$	$0.783\pm 0.012$	$0.936\pm 0.006$
2,400	$0.2069\pm 0.0044$	$0.2663\pm 0.0038$	$0.0191\pm 0.0223$	$0.797\pm 0.014$	$0.949\pm 0.006$
4,800	$0.1952\pm 0.0035$	$0.2540\pm 0.0037$	$-0.0579\pm 0.0166$	$0.811\pm 0.007$	$0.954\pm 0.005$

Table 6: Confirmation point error, Gaussian score, and empirical interval coverage for knowledge-context fusion. Values are mean

\pm

sample standard deviation across seeds.

Figure 4: Additional confirmation diagnostics. (a) Empirical coverage of the knowledge-context system across training sizes; error bars are standard errors over ten paired seeds and dashed lines mark nominal 80% and 95% coverage. (b) Prespecified encoder-level regularization contrasts with bootstrap 95% intervals. Each legend label is written as the first system minus the second, so negative aligned-minus-marginal values favor aligned regularization and positive wrong-minus-aligned values indicate harm from the wrong regularizer.

Figure 5: Confirmation audit of graded context contamination and structurally distinct prior errors. Effects are specific to the declared four-dimensional synthetic latent convention and do not test the validity of TCM theory.

Train $n$	Wrong excess MAE [95% CI]	AUROC	AP	Wrong flag rate	Damage $\rho$
600	$0.002410\ [0.002001,0.002795]$	0.620	0.608	0.107	0.014
1,200	$0.001999\ [0.001470,0.002514]$	0.614	0.606	0.112	0.022
2,400	$0.001823\ [0.001390,0.002213]$	0.612	0.603	0.109	0.031
4,800	$0.001519\ [0.001096,0.001913]$	0.610	0.599	0.105	0.045

Table 7: Wrong-context harm and descriptive conflict diagnostics. The threshold is the aligned-score 95th percentile;

\rho

is the seed-averaged Spearman association between score increase and case-level prediction damage.

Multimodal Reasoning

Multimodal reasoning studies how AI systems combine text with images, audio, video, and other signals to solve evidence-dependent tasks.

See all 512 papers

Oct 28, 2026

Auditable Hierarchical Inference with Multimodal Evidence and Bayesian Active Inquiry

Most multimodal classifiers assume that all useful evidence is available upfront, whereas clinical inquiry is sequential. We present a synthetic evaluation framework separating objective multimodal inference from active inquiry. An objective model maps synthetic tongue, bilateral multipressure pulse, and acoustic observations to eight syndrome families. A provenance-bearing Knowledge Core and training-only answer model support Bayesian narrowing over 33 fine terms using 25 typed questions, while seven safety questions remain outside the diagnostic posterior. Weak shared symptom factors limit fine-term influence on objective phenotypes, and an objective-only probe tests leakage. In a preregistered 600-case synthetic evaluation, objective family accuracy was 87.0%. Cost-sensitive expected information gain (cost-EIG) raised family accuracy to 95.5% and exact fine-term accuracy to 42.5%, versus 92.7% and 33.0% for an unoptimized fixed-order baseline. Selective family accuracy reached 96.9% at 95.3% coverage. However, cost-EIG asked more questions on average than fixed order (9.795 versus 9.352). Missing-audio routing retained 94.0% family accuracy, and a workflow invariant enforced abstention when all objective modalities were absent. Permuting fine-term answer likelihoods reduced family accuracy to 78.7%. Five of six frozen checks passed, but the question-efficiency hypothesis failed. Comparisons are descriptive because only aggregate, not paired case-level, outputs were retained. These results establish a reproducible synthetic workflow, not clinical diagnostic or treatment validity.

Yucheng Liu, Everett X. Wang

Department of Engineering Science, University of Oxford, Oxford, United Kingdom · Beijing Institute of Nanoenergy and Nanosystems, Chinese Academy of Sciences, Beijing, China

Multimodal Classification Clinical Decision-Making Multimodal Model Evaluation Bayesian Experimental Design Synthetic Data Evaluation

Ask PDF

Oct 17, 2026

TCMScience: An Autonomous Scientist for Traditional Chinese Medicine

Traditional Chinese medicine (TCM) research connects classical texts, medicinal entities, molecular assays, computational predictions, and clinical studies. An autonomous research agent must preserve the inferential boundaries between these evidence types while coordinating analysis across heterogeneous resources. We introduce TCMScience, a scientific knowledge-discovery agent that separates language-model proposals from execution and release authority. Its architecture combines a trusted kernel, a scientific compiler, and claim-support contracts: plans are compiled into typed operations, execution is bounded by explicit permissions, and candidate findings are checked against source provenance, study design, and claim scope before release. The full capability registry contains 117 public-source connectors with 444 typed operations across 29 domains and 150 native tools across 12 domains; the TCM data hub catalogs 136 databases, 76 dataset specifications, and 84,294 formula-source records. These resources support model-independent workflows through browser and local execution interfaces. In a network-pharmacology analysis of Gegen Qinlian Tang, replacing the whole-proteome background with the assayed-protein background changed the number of significant pathways from 66 to zero; incorporating inactive screening records identified nine pathways, including drug-metabolism signals retained as mechanistic hypotheses. A governance ablation over 121 constructed erroneous outputs spanning 26 classes released 121 errors without governance, 57 with the domain layer, seven with the kernel, and zero with both layers. All 10 compliant controls were retained. These results establish complementary control of known failure modes and show how explicit evidence contracts change the conclusions an agent is permitted to report.

Yanlan Kang, Ruiqi Liu, Xinpei Gu +4

Institute of Medical Philosophy & Future AI (IMPF-AI) · Shanghai Medical College, Fudan University · Guangdong Provincial Key Laboratory of New Drug Screening and Guangdong–Hong Kong–Macao Joint Laboratory for New Drug Screening, School of Pharmaceutical Sciences, Southern Medical University +4

AI Agents for Scientific Discovery Scientific Workflow Automation AI Agent Governance Scientific Claim Verification Traditional Chinese Medicine

Ask PDF

Oct 28, 2026

Fu-TCM: a multimodal large language model that learns traditional Chinese medicine diagnostic reasoning

Large language models (LLMs) are advancing in multimodal understanding and reasoning, accelerating medical model development. Traditional Chinese medicine (TCM) uses evidence from the four examinations to determine disease location and nature, differentiate syndromes and guide treatment. This diagnostic process constitutes a complete chain of reasoning. We present Fu-TCM, a multimodal TCM LLM trained on data retained after cleaning and filtering an initial collection of approximately 7.12 million examples derived from classical texts, textbooks, medical images, clinical cases and publicly available datasets. The model was further optimized using Bianzheng-Grounded Policy Optimization (BGPO), which explicitly targets syndrome differentiation through rewards for response format, intermediate TCM reasoning fields and final syndrome predictions. Fu-TCM-27B led all six benchmarks spanning TCM reasoning, text and vision, achieving a macro-average accuracy of 81.01% and surpassing the strongest reference model with complete results across all six benchmarks by 7.13 percentage points. Its largest margin was 11.90 points on Fu-bianzheng. Case-level analyses further showed a positive association between intermediate-field accuracy and final syndrome-selection accuracy in most evaluated models. These findings suggest that TCM syndrome differentiation may be computationally learnable and that structured reasoning supervision can improve performance across TCM tasks.

Jiahang Xie, Tong Jin, Lingfei Wu +9

School of Data Science, Fudan University, Shanghai, China · School of Life Sciences, Beijing University of Chinese Medicine, Beijing, China · The Affiliated People’s Hospital of Fujian University of Traditional Chinese Medicine, Fuzhou, Fujian, China +3

Medical LLMs Traditional Chinese Medicine MLLM Reasoning Structured Reasoning in Language Models RL for Language Model Reasoning

Ask PDF

Abstract

Multimodal Reasoning

Explore similar work

Auditable Hierarchical Inference with Multimodal Evidence and Bayesian Active Inquiry

TCMScience: An Autonomous Scientist for Traditional Chinese Medicine

Fu-TCM: a multimodal large language model that learns traditional Chinese medicine diagnostic reasoning