M-Reason: LLM-Based Biomedical Synthesis

• M-Reason is a multi-agent LLM-based system that synthesizes heterogeneous biomedical evidence with a focus on cancer research and transparent, audit-friendly workflows.
• It decomposes the evidence analysis into specialized modules for retrieval, analysis, and integration, ensuring structured outputs and deterministic orchestration.
• The system emphasizes reproducibility, traceability, and controlled synthesis by separating deterministic coding from LLM reasoning to minimize errors like hallucinations.

M-Reason is a multi-agent, LLM-based biomedical evidence synthesis system designed for transparent, auditable reasoning over heterogeneous biomedical evidence, with a particular emphasis on cancer research. It is presented as a demonstration platform and an interactive testbed rather than a purely offline benchmark system. Its central objective is to transform scattered evidence sources into a structured, question-specific synthesis by decomposing biomedical analysis into specialized evidence workflows and a higher-level integration process. The system is motivated by persistent limitations in biomedical LLM applications, including difficulty handling heterogeneous evidence sources, lack of provenance and traceability, hallucination risk, weak expert auditability, and the gap between research prototypes and robust user-facing systems (Wysocki et al., 6 Oct 2025).

1. Definition, scope, and intended use

M-Reason is oriented toward biomedical researchers, especially cancer researchers, and is integrated into a portal for numerical analysis and comparison of cancer patient samples. In this setting, analytical outputs can be passed directly into the system’s agents for downstream evidence synthesis. The paper frames typical use cases as interpreting gene lists arising from upstream computational analyses, reviewing clinical variant evidence, incorporating pharmacogenomic knowledge, interpreting gene enrichment results, and distinguishing well-known findings from potentially novel biomarkers or implications (Wysocki et al., 6 Oct 2025).

The system is explicitly designed as a workflowed reasoning environment rather than a direct-answer assistant. Instead of asking a single model to answer from broad context, M-Reason separates evidence retrieval, source-specific interpretation, cross-source synthesis, and report validation. This design reflects five priorities stated in the paper: interactive transparency, auditable reasoning, structured and reproducible synthesis, extensibility, and reliable scientific reporting. A plausible implication is that the system treats biomedical reasoning as a process of evidence management and controlled synthesis rather than as a single generative act.

The paper emphasizes that M-Reason is best understood as a structured, auditable biomedical reasoning workflow. It is not described as a free-form autonomous “agent society.” The authors explicitly note that it functions more as a deterministic workflow than a fully autonomous agent, a design choice intended to improve predictability and reproducibility even at the cost of reduced adaptive autonomy (Wysocki et al., 6 Oct 2025).

2. System architecture and evidence streams

The architecture is modular, layered, and two-stage. The first stage consists of independent evidence retrieval and assessment modules. The second stage consists of a higher-level evidence integration and report synthesis layer. The source-specific evidence streams named in the paper are CIViC for clinical variant evidence, PharmGKB for pharmacogenomic information, and g:Profiler for gene enrichment analysis (Wysocki et al., 6 Oct 2025).

Each source-specific module follows a three-role pattern: Orchestrator, BioExpert, and Evaluator. After these modules complete, a separate five-agent integration system operates with an Orchestrator, ReportComposer, ContentValidator, CriticalReviewer, and RelevanceValidator. This division enforces separation between source-level evidence appraisal and cross-source report synthesis (Wysocki et al., 6 Oct 2025).

A central architectural distinction is the separation between deterministic control logic and LLM-based reasoning. The Orchestrator in the evidence analysis system does not use an LLM. The Orchestrator in the integration system also operates via Python logic alone, parsing upstream outputs, consolidating evidence, dispatching tasks, and tracking workflow status. By contrast, BioExpert, Evaluator, ReportComposer, ContentValidator, CriticalReviewer, and RelevanceValidator are LLM-based agents responsible for content generation, review, and synthesis (Wysocki et al., 6 Oct 2025).

State is maintained primarily through explicit workflow artifacts rather than latent conversational memory. The paper describes structured JSON outputs at each stage, iteration history, consolidated evidence objects passed into the integration phase, logged prompts and intermediate outputs, and workflow status tracked by orchestrators. This suggests an architecture in which serialized intermediate representations are the primary memory substrate.

The end-to-end data flow is described procedurally. A user provides analysis context, a research question, and genes or uploaded JSON data. Source-specific modules retrieve evidence from databases or APIs. Within each module, BioExpert analyzes the evidence and Evaluator reviews the result, with revision loops up to a maximum limit. The integration orchestrator then parses all approved outputs, consolidates them into structured objects, and passes them to the integration layer. ReportComposer synthesizes a unified report, three reviewers evaluate it in parallel, and if unanimous approval is not achieved, combined feedback is sent back for revision. Final outputs are then presented and can be exported as PDF, while logs, prompts, intermediate outputs, metrics, and iterations remain inspectable (Wysocki et al., 6 Oct 2025).

3. Agent roles and reasoning workflow

The Orchestrator is purely procedural. In the source-specific pipeline, it initiates the workflow, delegates tasks to BioExpert and Evaluator, tracks iteration count and status, and enforces stopping conditions. The paper explicitly states that this agent does not perform reasoning or content analysis and does not interact with any LLM (Wysocki et al., 6 Oct 2025).

BioExpert is the primary source-specific analysis agent. It analyzes retrieved evidence in light of the user’s research context and question, then produces structured output with relevance explanations, summaries, conclusions, and explicit citations to evidence sources. The specialization of BioExpert is defined by evidence stream rather than model class: one instance may analyze CIViC evidence, another PharmGKB evidence, and another enrichment results. The prompting scheme changes across iterations. On the first pass, BioExpert receives detailed role instructions plus context, question, and evidence. On later passes, it also receives the previous analysis and Evaluator feedback, and is instructed to revise while preserving valid prior content (Wysocki et al., 6 Oct 2025).

Evaluator is the quality-control agent at the source level. It reviews BioExpert’s output against the research context, question, evidence, and quality criteria, and must begin its response with either “APPROVED” or “NOT APPROVED.” If the answer is not approved, it must provide actionable bulleted feedback. The review criteria include scientific accuracy, citation quality, clarity, and completeness. Communication is mediated by the Orchestrator rather than through free-form inter-agent dialogue (Wysocki et al., 6 Oct 2025).

At the integration layer, the integration Orchestrator is again deterministic and Python-based. ReportComposer performs the cross-source synthesis. It is required to use only the provided evidence, use bullet-point formatting, include evidence citations, and organize the report into four mandatory sections: potential novel biomarkers, implications, well-known interactions, and conclusions (Wysocki et al., 6 Oct 2025).

The three integration reviewers have distinct roles. ContentValidator checks formal structure, evidence grounding, citation presence, coverage of evidence sources, and whether unsupported external information was introduced. CriticalReviewer performs adversarial scientific review, looking for bias, unsupported claims, and alternative interpretations. RelevanceValidator checks whether the report answers the research question, whether novelty versus well-known findings are properly classified, and whether conclusions are logically supported. Their outputs are combined through a unanimous consensus rule: if all approve, the report is accepted; otherwise, collective feedback is returned to ReportComposer for revision (Wysocki et al., 6 Oct 2025).

This layered review protocol is the system’s main mechanism for converting agent specialization into auditable synthesis. A plausible implication is that M-Reason treats disagreement as a signal for controlled refinement rather than as an error to suppress.

4. Explainability, auditability, and deterministic validation

Explainability and auditability are the central design goals of M-Reason. The system requires explicit source citations in generated analyses and final reports, including direct links when available. The paper states that synthesized outputs are fully traceable to source evidence. This traceability is supported by source-specific outputs, citations embedded in analyses, structured JSON artifacts, final report sections linked back to upstream evidence, and exposed prompts and logs (Wysocki et al., 6 Oct 2025).

The user can inspect all evidence under consideration, each agent’s communications and outputs, intermediate analyses, reviewer feedback, all prompts, logs from each iteration, execution metrics, and the final report. This is unusually strong process exposure relative to black-box LLM assistants and is integral to the claim of user auditability (Wysocki et al., 6 Oct 2025).

Structured outputs are also central. The paper states that all outputs at every stage are generated in structured JSON format, although the exact schema is not reproduced. The final integration report has a fixed organization with the four mandatory sections noted above. This structure serves both usability and machine auditability. It also makes module-to-module handoff easier and reduces variability in downstream integration (Wysocki et al., 6 Oct 2025).

Deterministic validation enters primarily through orchestration and workflow control rather than formal symbolic verification. The two Orchestrators are non-LLM components, and the integration Orchestrator operates through Python logic alone. These components handle sequencing, file parsing, evidence consolidation, status tracking, dispatch control, and loop termination. The paper repeatedly stresses the value of combining LLM reasoning with deterministic code to improve confidence, predictability, and reproducibility (Wysocki et al., 6 Oct 2025).

The most precise formal rules in the system are procedural rather than mathematical. At the source level, Evaluator must respond with either APPROVED or NOT APPROVED plus actionable feedback. At the integration level, acceptance requires unanimous approval among ContentValidator, CriticalReviewer, and RelevanceValidator. The paper does not provide explicit equations for evidence scoring, confidence aggregation, novelty scoring, reviewer weighting, or consensus optimization (Wysocki et al., 6 Oct 2025).

Because outputs, prompts, and workflow state are serialized and logged, reproducibility depends less on hidden context accumulation and more on explicit artifact preservation. This design suggests a conception of explainability rooted in provenance and inspectable intermediate products rather than in post hoc natural-language justification alone.

5. User interface, interaction model, and evaluation

The interface is a major part of the contribution. Users can specify the overall analysis context, the specific research question, and a list of genes. Gene input can be manual, imported from other analyses within the cancer-sample portal, or uploaded as JSON. The interface is therefore tightly coupled to upstream computational workflows rather than being an isolated chat system (Wysocki et al., 6 Oct 2025).

A distinctive UI feature is the display of the live multi-agent workflow through seven distinct terminals, each showing the activities and communications of an individual agent. The paper describes this as enabling users to observe ongoing interpretation of evidence, structured report generation, and parallel feedback from validation agents. Comprehensive evidence tracking is explicitly emphasized: users can monitor all evidence under consideration, follow how it is processed, and observe live parallel execution (Wysocki et al., 6 Oct 2025).

Upon completion, the final report is shown in the interface and can be downloaded as PDF. The UI also surfaces operational metadata, including processing time, computational cost, token usage, number of genes analyzed, and number of iterations. This directly exposes tradeoffs between quality control and resource usage (Wysocki et al., 6 Oct 2025).

The evaluation is deliberately lightweight and focuses on feasibility, efficiency, and consistency. The authors test four cumulative gene-list scenarios: $S_1$ with 13 genes, $S_2$ with 28 genes, $S_3$ with 52 genes, and $S_4$ with 82 genes. These correspond to evidence sets ranging from 1,656 words to 81,627 words. For consistency, five independent executions were run for each scenario with the same context, question, and gene list, and outputs were reviewed by human experts and an LLM (Wysocki et al., 6 Oct 2025).

The baseline for time efficiency is a human expert reading speed of 200 words per minute. The paper explicitly notes that this baseline estimates time to read the evidence, not to analyze it or write a report. Within that framing, the strongest quantitative claim is that in the largest scenario $S_4$, M-Reason generated a report approximately 135 times faster than a human would need merely to read the evidence (Wysocki et al., 6 Oct 2025).

Qualitatively, the paper reports that M-Reason preserved all critical findings as evidence size increased and that repeated runs consistently highlighted the same novel and well-known genes. This suggests output stability under repeated execution, although no formal variance statistic is given. A plausible implication is that structured workflow control and consensus review reduce run-to-run drift relative to unconstrained single-agent generation.

6. Tradeoffs, limitations, and future directions

The paper explicitly identifies a tradeoff between agent specialization and system complexity. More specialized agents may improve rigor and accuracy, but they also increase orchestration overhead, maintenance burden, and resource usage. This is one of the system’s central design tensions (Wysocki et al., 6 Oct 2025).

A second tradeoff concerns the combination of LLM reasoning with deterministic code. The authors argue that this improves confidence, predictability, and reproducibility, but can reduce speed, flexibility, and adaptive autonomy. This is directly connected to their statement that M-Reason operates more as a deterministic workflow than a fully autonomous agent (Wysocki et al., 6 Oct 2025).

Resource cost is also a recurring concern. Because the system uses multiple source-specific pipelines, iterative review loops, parallel reviewers, and integration-stage revision, token usage, runtime, and API cost are nontrivial. The interface exposes these metrics precisely because they matter operationally. The reported evaluations use GPT-4.1-mini throughout, so system behavior and cost may differ with other models (Wysocki et al., 6 Oct 2025).

Hallucination remains a residual risk. The system attempts to reduce it through source-bounded prompts, anti-hallucination instructions, citation requirements, evaluator review, content validation, and consensus-based acceptance. However, the paper does not claim hallucination is eliminated, and core evaluators are themselves LLMs. The current evidence coverage is also limited to three sources—CIViC, PharmGKB, and gene enrichment via g:Profiler—which constrains breadth and domain coverage (Wysocki et al., 6 Oct 2025).

The evaluation is intentionally modest. The paper does not report direct comparisons against single-agent systems, ablations removing specific reviewers, inter-rater agreement, detailed latency breakdowns by module, or formal factuality metrics. Its empirical contribution is therefore strongest as a demonstration of feasibility, transparency, and consistency rather than as a definitive benchmark victory.

Future work is framed around expanding evidence-source coverage, especially PubMed integration, improving evaluation rigor, balancing specialization against cost, strengthening deterministic validation, testing other model backbones, and possibly increasing adaptive autonomy while preserving auditability (Wysocki et al., 6 Oct 2025).

Taken together, these limitations clarify the paper’s positioning. M-Reason is not presented as a complete solution to biomedical evidence synthesis. It is a practical and inspectable architecture showing how multi-agent LLM systems can be made more trustworthy for scientific use by combining source-specific specialization, deterministic orchestration, iterative review, structured outputs, and user-visible provenance (Wysocki et al., 6 Oct 2025).