Paper deep dive
AI Economist Agent: An Agentic Framework for Model-Grounded Economic Analysis with RAG, Knowledge Graphs, and Large Language Models
Masahiro Kato
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 6/21/2026, 5:00:10 AM
Summary
The paper proposes the 'AI Economist Agent,' an agentic framework designed for model-grounded economic analysis. Unlike standard LLM-based RAG which focuses on text retrieval, this framework utilizes a Knowledge Graph (KG) to integrate economic theory, text evidence, and mathematical model specifications. The architecture employs LLM-based agents to plan analysis, retrieve context, and select specific mathematical models (e.g., DSGE-lite, VAR) for execution. This ensures that economic claims in generated reports are grounded in explicit model-based computations rather than just linguistic fluency. The framework was evaluated through two applications: U.S. inflation persistence/Federal Reserve policy and U.S. commercial real estate (CRE) refinancing stress, demonstrating that the model-grounded GraphRAG approach improves economic coherence and traceability compared to LLM-only or RAG-only methods.
Entities (10)
Relation Signals (7)
Knowledge Graph → contains → Model Specification
confidence 100% · The knowledge graph contains text-style economic reports, temporal facts, model specifications and mathematical model outputs.
AI Economist Agent → evaluatedon → U.S. Commercial Real Estate Refinancing Stress
confidence 100% · and bank stress-test narrative generation for U.S. commercial real estate (CRE) refinancing stress.
AI Economist Agent → evaluatedon → U.S. Inflation Persistence
confidence 100% · We evaluate the AI economist agent in two applications: economist report generation for U.S. inflation persistence and Federal Reserve policy
AI Economist Agent → uses → Large Language Models
confidence 100% · utilizes knowledge graphs including economic data and theory and LLM-based agents
AI Economist Agent → utilizes → Knowledge Graph
confidence 100% · utilizes knowledge graphs including economic data and theory
Planner Agent → generates → Analysis Plan
confidence 90% · The planner agent first reads the scenario and produces an analysis plan.
Model-Selection Agent → selects → Model Specification
confidence 90% · The model-selection agent receives the scenario, the plan, and the graph-retrieved model context.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We propose a model-grounded RAG-based AI economist with an agentic framework for economic scenario analysis using large language models (LLMs) and knowledge graphs. While LLMs can generate fluent economic narratives, economists are often required to make economic claims grounded by economic theory and real-world data. Based on this motivation, this study proposes an RAG-based AI economist, which utilizes knowledge graphs including economic data and theory and LLM-based agents to plan the analysis, retrieve relevant evidence, select appropriate models, and generate reports. In our framework, we do not produce quantitative claims directly with the language model alone; instead, we generate narratives grounded in explicit model-based computations and linked to the retrieved evidence via AI agents. We refer to our framework as an AI economist agent. We evaluate the AI economist agent in two applications: economist report generation for U.S. inflation persistence and Federal Reserve policy, and bank stress-test narrative generation for U.S. commercial real estate refinancing stress. The results illustrate how grounding the generated reports improves their economic coherence and traceability.
Tags
Links
- Source: https://arxiv.org/abs/2606.20041v1
- Canonical: https://arxiv.org/abs/2606.20041v1
Trouble viewing inline? Open PDF directly →
Full Text
46,055 characters extracted from source content.
Expand or collapse full text
AI Economist Agent: An Agentic Framework for Model-Grounded Economic Analysis with RAG, Knowledge Graphs, and Large Language Models Masahiro Kato ∗ Mizuho-DL Financial Technology, Co., Ltd. June 19, 2026 Abstract We propose a model-grounded RAG-based AI economist with an agentic framework for economic scenario analysis using large language models (LLMs) and knowledge graphs. While LLMs can generate fluent economic narratives, economists are often required to make economic claims grounded by economic theory and real-world data. Based on this motivation, this study proposes an RAG-based AI economist, which utilizes knowledge graphs including economic data and theory and LLM-based agents to plan the analysis, retrieve relevant evidence, select appropriate models, and generate reports. In our framework, we do not produce quantitative claims directly with the language model alone; instead, we generate narratives grounded in explicit model-based computations and linked to the retrieved evidence via AI agents. We refer to our framework as an AI economist agent. We evaluate the AI economist agent in two applications: economist report generation for U.S. inflation persistence and Federal Reserve policy, and bank stress-test narrative generation for U.S. commercial real estate refinancing stress. The results illustrate how grounding the generated reports improves their economic coherence and traceability. Keywords: AI economist; large language models; retrieval-augmented generation; GraphRAG; knowledge graphs; economic report generation; macro economics; finance; bank stress testing 1 Introduction Artificial intelligence is increasingly used in economic analysis, and large language models (LLMs) have drawn attention as tools for generating economic reports. However, economic analysis is not only a language-generation issue. A statement such as “persistent inflation will ∗ Email: mkato-csecon@g.ecc.u-tokyo.ac.jp 1 arXiv:2606.20041v1 [econ.GN] 18 Jun 2026 keep monetary policy restrictive” becomes economically meaningful when it can be traced to a macroeconomic mechanism, relevant data, and an explicit model. This distinction is central to the design of AI systems in economics. Economists often move between narrative and formal reasoning. In inflation analysis, a narrative about persistence must be connected to price-setting behavior, labor market conditions, and monetary policy reactions. In financial stress analysis, a narrative about risk must explain how macro-financial shocks pass through asset prices, funding conditions, bank losses, capital, and liquidity. Fluent text generation alone is insufficient for these tasks. We propose a model-grounded RAG-based AI economist. The framework uses AI agents with retrieval-augmented generation (RAG) on knowledge graphs. The knowledge graph contains text-style economic reports, temporal facts, model specifications and mathematical model outputs. Agents operate on this graph so that a report can be traced from the scenario to the relevant evidence and an executed model. They plan the analysis, retrieve context, select models, and generate reports, but do not generate quantitative model-dependent claims on their own. These claims are based on computations performed using mathematical models. We examine the framework in two applications. The first is economist report generation for U.S. inflation persistence and Federal Reserve policy. The second is the bank stress-test narrative generation for U.S. commercial real estate (CRE) refinancing stress. For each application, we compared three report-generation conditions: LLM-only, RAG-only, and model-grounded GraphRAG. The LLM-only condition received only the scenario narratives. The RAG-only condition received the scenario narrative and retrieved text evidence. The model-grounded GraphRAG condition adds graph paths and a selected model specification and then uses the resulting model run and model-output summaries. Contribution. This study contributes to the agentic architecture for economic report generation. It runs AI agents over a knowledge graph that contains both text-style information and mathematical methods and grounds model-dependent economic claims in executed mathematical models. Much of the current use of LLMs in economic writing emphasizes fluent, coherent text generation. Here, the LLM is treated as an agent that plans, retrieves, selects a model, and explains the result, whereas the numerical component is produced by explicit model execution. We refer to this architecture as an AI economist agent. This terminology is related to, but distinct from, the AI Economist of Zheng et al. (2022), which addresses a different economic-design problem. The architecture is intended for settings in which economic narratives must be linked to evidence and model outputs, including economist reports and stress test documentation. 2 Related Work This section positions the study within four strands of research. AI agents for economic research. In economics, AI economist frameworks have been proposed to support literature reviews, data retrieval, computational assistance, and ex- ploratory analyses Korinek (2023, 2025). This study specializes in the agentic idea of report 2 generation in economics. The proposed AI economist agent extends this line of work by connecting text generation to mathematical models. Macroeconomic knowledge graphs. Existing studies on macroeconomic knowledge graphs have used economic documents and reports to represent the relationships among macroeconomic variables and alternative data variables, often for variable selection in fore- casting models Yang et al. (2020). Our framework extends this logic by adding mathematical models to the graph schema. Therefore, the graph is used not only to organize documents but also to connect scenarios, evidence, model specifications, and model runs. RAG and GraphRAG. RAG retrieves text spans, whereas GraphRAG retrieves structured paths. Agentic RAG and graph-based RAG systems have been proposed for retrieval with planning, graph traversal, and multi-step reasoning Singh et al. (2026); Guo et al. (2025); Gutierrez et al. (2024); Liang et al. (2025); Zhang et al. (2025). Broad graph expansion can also retrieve excessive context, especially when a task requires a narrow answer Chen et al. (2025). Therefore, this study uses task-specific graph queries rather than unconstrained neighborhood expansion. Model grounding and temporal graph design. Work on mathematical reasoning and external computation shows that tool-based calculations can ground LLM reasoning Jung et al. (2025). The economic analog is to execute approved macro-financial model functions rather than ask an LLM to invent numerical paths. Economic analysis is also time-sensitive; the date of a document, validity period of a claim, and execution time of a model run should be represented separately. Dynamic temporal knowledge graph construction and agent memory systems motivate this separation Lairgi et al. (2026). This study adapts these ideas to a controlled setting in which evidence spans, temporal facts, model specifications, model runs, and outputs are represented separately in the graph. 3 AI Economist Agent The AI economist agent is designed to generate economic reports whose statements can be traced to structured knowledge and mathematical analysis. The LLM plans the analysis, controls retrieval, selects models, and writes reports. It does not generate quantitative outputs by itself. Numerical paths and stress metrics are produced by computational functions selected through the knowledge graph. 3.1 Overview of the Architecture The AI economist agent has three main features. First, text evidence and mathematical models are stored as different objects. Economist-style documents provide evidence for narrative statements, whereas model specifications define what can be computed. Second, the graph is used not only for text retrieval but also for model selection. The model-selection agent receives model specifications and implementation mappings retrieved from the graph. 3 Third, model-dependent report content is generated only after an executable model has been selected and run. Letsdenote a scenario. Text retrieval returns evidence spansE s associated withs. Graph retrieval before model execution returns pathsP pre s . The graph also returns executable model contextC s , which contains model specifications and implementation mappings. The model-selection agent produces a structured requestr s . Then, the computational layer executes the model specified inr s , produces output pointsO s , and records a model runm s in the graph. A second graph query returns post-run pathsP post s . The proposed report is generated from (s, E s , P pre s , C s , r s , m s , O s , P post s ). This sequence is the main difference between ordinary RAG-only report generation and the proposed model-grounded GraphRAG condition. 3.2 Knowledge Graph The knowledge graph is the organizing structure of the system. It is a directed graph with typed nodes and typed relationships. Its role is not to store all economic knowledge as natural language text. Instead, it separates objects that are often conflated in LLM-based report generation: text evidence, economic concepts, model specifications, and model execution results. Table 1 summarizes the main objects. The table groups node labels by their function in the architecture. Text-evidence objects are used by the RAG-only condition. Temporal objects represent observation-time and validity-window information attached to evidence relations. Model-catalog objects are used by the proposed condition when the model-selection agent chooses an executable specification. Execution objects are created after the selected model is run. Table 1: Main graph objects. Object groupNode labelsPurpose Text evidence SourceDocument, EvidenceSpanProvides retrieved evidence for RAG-only report generation. Temporal evidence TemporalFact Records observation time and validity-window information. Economic task Scenario,Shock, VariableConcept Represents the scenario and affected variables. Model catalog Model,ModelSpecification, ModelEquation, ModelAssumption, ModelDocument Stores approved model descriptions separately from text evidence. Execution ImplementationFunction , ModelRun,ForecastOutcome, ModelOutputPoint, BankMetric Records executable mappings and computed outputs. 4 The separation between text evidence and model specifications is essential. A text span may state that inflation persistence is associated with a restrictive policy response. That statement is useful evidence, but it is not itself a model. A model specification is stored as a separate object with equations, assumptions, inputs, outputs, and an implementation mapping. This design prevents the RAG-only condition from becoming a model-selection condition simply by reading model names from retrieved text. The graph also records model execution. When a model is executed, the system creates a ModelRunnode linked to the scenario and the selected model specification. The run produces summary objects and horizon-level output points. If the application involves bank stress-test quantities, the run also producesBankMetricnodes. The post-run graph therefore contains a path from the scenario to the model specification, implementation function, model run, and computed outputs. 3.3 Model Specifications and Executable Implementations A model specification and an executable implementation have different roles. A model specification describes the economic or statistical model that is appropriate for a scenario. It may include equations, assumptions, input variables, output variables, and model-governance text. An executable implementation is the computational family that computes a numerical path. Several model specifications may map to the same implementation family. This distinction is useful for two reasons. First, the graph can contain many approved model specifications without requiring a separate executable routine for each one. Second, the LLM does not select an arbitrary model name from free text. It selects a model object returned by graph retrieval and maps it to an available implementation family. The resulting request has a controlled structure: r s = (model specification, model family, implementation family, execution tool, expected outputs). The computational runner validates this request against the implementation mapping and then executes the corresponding model function. This design makes the proposed condition model-grounded in a narrow sense. The LLM chooses among graph-retrieved model objects and explains the choice. The numerical path is computed outside the LLM. The generated report can therefore distinguish between narrative interpretation and model output. 3.4 Agent Workflow Each application is processed by the same workflow. The graph is cleared at the beginning of the run and then loaded with the evidence rows, temporal facts, model catalog, scenario node, and scenario-shock relationships. This reset is important because the post-run graph should reflect the current application rather than previous graph contents. The planner agent first reads the scenario and produces an analysis plan. The plan specifies the economic objective, target variables, relevant shocks, and model-selection objective. The text-retrieval step then returns evidence spans for the RAG-only and proposed conditions. In 5 parallel, the graph-retrieval step returns task-specific paths from the knowledge graph. These paths expose the scenario-variable-evidence structure and retrieve executable model context. The model-selection agent receives the scenario, the plan, and the graph-retrieved model context. It returns a structured model-execution request rather than a free-form recom- mendation. The computational layer executes the requested implementation family and produces a model-output table. The system writes the resultingModelRun,ForecastOutcome, ModelOutputPoint, andBankMetricnodes back to the graph. After this write-back, the graph is queried again. The proposed report is generated only after this second query. The final step generates reports under the three conditions and evaluates them. The judge agent scores each report using the same criteria, but the conditions receive different inputs. The comparison is therefore not a test of whether a longer prompt produces a longer report. It asks whether graph retrieval and model execution change the grounding structure of the generated report. 3.5 Report-Generation Conditions The architecture compares three report-generation conditions. Table 2 summarizes the information available to the report generator. The LLM-only condition receives only the scenario narrative. The RAG-only condition receives the scenario narrative and retrieved text evidence. The model-grounded GraphRAG condition receives graph paths and model- execution results in addition to retrieved evidence. Table 2: Report-generation conditions. ConditionInput to the report generatorIntended comparison LLM-onlyScenario narrative.Baseline narrative without external grounding. RAG-onlyScenario narrative and retrieved text evidence. Textual grounding without model execution. Model-grounded GraphRAG Scenario narrative, retrieved evi- dence, graph paths, selected model specification, model run, and model outputs. Proposed condition with graph retrieval and executed models. 4 Experiments The experiments check whether the information reaching the report generator is controlled as intended. Application 1 and Application 2 run the same architecture under two economic narratives. The comparison is therefore not an evaluation of the substantive economic content of the two applications; it is an inspection of the report-generation process. In the LLM-only condition, the report generator receives only the scenario. In the RAG-only condition, it also receives retrieved text evidence. In the model-grounded GraphRAG condition, it receives the retrieved evidence, graph context, and the results of model execution. 6 Prompt design and model execution. The prompts are condition-specific. The LLM- only prompt asks for an economist-style narrative based only on the scenario. The RAG-only prompt asks the model to use retrieved evidence and not to claim that a mathematical model has been executed. The proposed prompt gives the report generator the selected model specification, the execution request, the model-run identifier, the output summary, and the post-run graph paths. It asks the report generator to distinguish qualitative interpretation from model-dependent statements. The three conditions differ by construction. RAG-only generation can improve textual grounding because it supplies retrieved evidence. It cannot provide model grounding unless a model run is supplied. The proposed condition adds this missing component by connecting the report to a graph-selected model specification, an executed implementation, and output nodes written back to the knowledge graph. Graph construction. The graph is rebuilt for each application. At the beginning of the run, the existing graph is cleared. The system then loads economist-report evidence, temporal facts, the model catalog, the scenario node, and the scenario-shock relationships. This reset ensures that the post-run graph is specific to the current application. It also makes the path from the scenario to the model output inspectable after model execution. The model catalog is not embedded in the evidence prose supplied to the RAG-only report generator. Evidence spans are used by RAG-only generation as textual support. Model specifications are stored separately as graph objects and are linked to executable implementation families. This separation is important because the RAG-only condition should not become a model-selection condition by reading model recommendations from retrieved text. Table 3 reports the size of the graph inputs used in the run. Table 3: Knowledge-graph inputs. ComponentCount Economist-report evidence rows2500 Temporal-fact rows2500 Executable implementation families14 ModelSpecification rows420 ModelEquation rows1260 ModelAssumption rows1260 ModelDocument rows420 Model-variable relationship rows2280 Table 4 summarizes the executable implementation families. The graph can contain many model specifications because several specifications may map to the same implementation family. The LLM selects a graph-retrieved specification and returns a structured execution request. The computational layer then validates the requested implementation family and computes the output path. The numerical path is therefore produced outside the LLM. The same prompt design is used in both applications. The LLM-only prompt asks the model to write from the scenario alone. The RAG-only prompt adds retrieved evidence and 7 Table 4: Executable model families. Implementation familyRole in the run VAR and BVARReduced-form macro-financial propagation. SVAR and Local ProjectionShock propagation and horizon-specific responses. DSGE-liteInflation, policy-rate, yield, and output-gap dynamics. Regime SwitchingNormal and stress dynamics. Stochastic VolatilityVolatility amplification. ARIMA and SARIMAXTime-series projection with or without exogenous inputs. State SpaceTrend-cycle projection. Factor-AugmentedVAR and GVAR Proxy Factor-based and spillover propagation. ML Ridge and Random- Forest Surrogate Predictive response functions over lagged variables. prohibits claims that a model has been executed. The proposed prompt receives the selected model specification, the execution request, the model-run object, the model-output table, and the post-run graph paths. It asks the report generator to separate economic interpretation from model-output information. Table 5 summarizes this design. Table 5: Information available to each report-generation condition. ConditionInputsWhat the condition can show LLM-onlyScenario narrative.Whether a fluent report can be written without external grounding. RAG-onlyScenario narrative and retrieved evidence spans. Whether textual evidence changes the report while model execution is absent. Model-grounded GraphRAG Scenario narrative, retrieved evidence, graph paths, selected model specification, model run, and model outputs. Whether the report can be connected to graph retrieval and an executed model. The same local LLM setting is used for planning, model selection, report generation, and judging. The graph store returns path results before and after model execution. The graph figures display selected returned paths after model execution. Labels are shortened and layout coordinates assigned for readability, but no economic nodes or relationships are added. Node markers are drawn as fixed-size circles. Model-output values are reported in synthetic macro-financial units over a 12-step horizon. 8 4.1 Application 1: Economist Report Generation Scenario and retrieved evidence. Application 1 concerns U.S. inflation persistence and Federal Reserve policy. The scenario states that services inflation remains persistent while labor-market cooling is gradual. The Federal Reserve keeps policy restrictive, which lifts Treasury yields and slows real activity. The application is used to inspect whether the proposed condition connects this narrative to retrieved evidence, a graph-selected DSGE-lite specification, and an executed model path. Table 6 reports the top retrieved evidence spans. The retrieved text is ranked by tag overlap with the scenario and then by the synthetic confidence score attached to each evidence span. As a result, the top set contains one directly monetary-policy-related span as well as broader macro-financial spans that share the scenario theme. This pattern is part of the comparison: RAG-only generation receives retrieved text as evidence, while the proposed condition separately queries the graph for model specifications and execution links. Table 6: Application 1 retrieved evidence. EvidenceMain statementConfidenceScore EV1161Credit losses reduce common equity capital and lower the stressed CET1 ratio. 0.9402.0 EV1321Declines in CRE values raise loss-given-default and can increase credit losses. 0.9392.0 EV 1801China growth surprises transmit to global trade volumes through supply-chain and commodity channels. 0.9362.0 EV1571A more restrictive policy path usually lifts nominal yields, although the term premium may vary by regime. 0.9352.0 EV841Global oil supply shocks tend to raise headline inflation before the pass-through fades. 0.9332.0 EV721Funding stress and confidence effects can increase uninsured deposit outflows. 0.9322.0 In Tables 6 and 10, “Confidence” is the synthetic confidence value stored with the evidence span, and “Score” is the retrieval score used to rank text evidence. It is not the judge score reported later. Report outputs. Table 7 summarizes the generated reports. The generated report texts are included in Appendix A. The LLM-only report is a scenario-based narrative. The RAG- only report cites retrieved evidence. The model-grounded GraphRAG report is linked to the selected DSGE-lite specification and to the executed model path. Table 7: Application 1 report comparison. ConditionReport content LLM-onlyNarrative based only on the scenario. RAG-onlyNarrative that cites EV1161, EV1321, EV1801, EV1571, EV841, and EV721. Model-grounded GraphRAG Narrative linked toMSPEC_DSGE_LITE_004, the model-run record, and the executed DSGE-lite output path. 9 Graph retrieval and model execution. Figure 1 displays selected paths from the post- ModelRunGraphRAG result. The path connects the scenario to shocks and variables, then to evidence and the selected DSGE-lite model specification. It also includes the implementation function, the model run, and output-point nodes. Figure 2 plots the executed model path. Table 8 reports the maximum, minimum, and final values over the 12-step synthetic path. These values are outputs of the experimental model layer and should not be read as empirical forecasts. A′#!*!&% &)*&#+%(' A'* ) C A′#!*!&% )%(!& &(!%#*!&% ) &" &#!,(* ) &" %,(,!# ) &" &( !%#*!&% +%) (* *%,( ,!# '&#!,,!# !%#*!&% #!* )'!!*!&% #!* !$'#$%**!&% &#+% #!% !%#*!&%'* +%) (*'* %,( ,!#'* +*'+*' '* Figure 1: Application 1 post-ModelRunGraphRAG paths. The figure displays selected paths from the returned graph result after model execution. Nodes are rendered as circular markers. Table 8: Application 1 model-output summary. VariablePeakTroughFinal U.S. headline inflation0.229-0.455-0.132 Federal funds rate30.951-5.144-1.130 U.S. ten-year yield17.1692.7314.337 U.S. output gap0.007-1.3000.007 Judged comparison. Table 9 and Figure 3 report the judge scores. The proposed condition receives higher graph and model-grounding scores than LLM-only and RAG-only. RAG-only receives a high evidence score because it cites retrieved evidence, but it does not receive a 10 1 1111 "$%&"$*"! 1 1 " "'&#'&'!&% ##&"!('&& "#& !!&"! $'!%$& &!)$) "'&#'&# Figure 2: Application 1 executed DSGE-lite model path. selected model specification or model output. The coherence score is not the main object of the comparison. The method is intended to change the grounding structure of model-dependent statements, not only to produce smoother prose. Table 9: Application 1 judge scores. ConditionEvidenceGraphModelNumericCoherenceAudit LLM-only2.01.00.05.04.02.0 RAG-only5.01.00.05.05.05.0 Model-grounded GraphRAG5.04.05.05.04.05.0 4.2 Application 2: Bank Stress-Test Narrative Generation Scenario and retrieved evidence. Application 2 concerns U.S. commercial real estate refinancing stress. Higher long-term rates and wider credit spreads put pressure on refinancing conditions. The scenario then passes through bank losses, funding conditions, capital, and liquidity. This application is used to inspect whether the same architecture can connect a stress-test narrative to a regime-switching model and to bank stress metrics. Table 10 reports the retrieved evidence. The spans focus on funding stress, uninsured deposit outflows, and liquidity coverage. The evidence therefore provides text grounding for RAG-only generation. The proposed condition adds graph paths and model execution to this text-grounded baseline. Report outputs. Table 11 summarizes the reports. The generated report texts are included in Appendix B. The LLM-only report discusses refinancing pressure without external 11 L& $# !$ ! ! #$"!"&# % ! $! ! $# Figure 3: Application 1 judge scores by condition. Table 10: Application 2 retrieved evidence. EvidenceMain statementConfidenceScore EV1967Funding stress and confidence effects can increase uninsured deposit outflows. 0.9392.0 EV2028Deposit outflows increase net cash outflows and reduce liquidity coverage ratios. 0.9382.0 EV2107Funding stress and confidence effects can increase uninsured deposit outflows. 0.9372.0 EV1197Funding stress and confidence effects can increase uninsured deposit outflows. 0.9312.0 EV628Deposit outflows increase net cash outflows and reduce liquidity coverage ratios. 0.9282.0 EV1267Funding stress and confidence effects can increase uninsured deposit outflows. 0.9262.0 12 grounding. The RAG-only report cites retrieved evidence. The model-grounded GraphRAG report is linked toMSPEC_REGIME_SWITCHING_005, the executed regime-switching path, and the bank stress metrics. Table 11: Application 2 report comparison. ConditionReport content LLM-onlyNarrative based only on the scenario. RAG-onlyNarrative that cites EV 1967, EV2028, EV2107, EV1197, and EV1267. Model-grounded GraphRAG Narrative linked toMSPEC_REGIME_SWITCHING_005, the executed model path, and bank stress metrics. Graph retrieval, model output, and stress metrics. Figure 4 displays selected paths from the post-ModelRunGraphRAG result. The path links the scenario to refinancing and funding variables, then to evidence and the selected regime-switching specification. It also shows the implementation function, model run, output points, and bank metrics. Figure 5 reports the executed model path. Table 12 reports the maximum, minimum, and final values over the 12-step synthetic path, and Table 13 reports the stress metrics. Figure 6 plots those metrics. Table 12: Application 2 model-output summary. VariablePeakTroughFinal CRE price-5.469-16.824-5.469 Credit spread152.97171.98271.982 Bank funding spread42.76030.59830.598 Deposit outflow rate4.3382.6982.698 Table 13: Application 2 bank stress metrics. MetricValue Stress PD0.049 Stress LGD0.467 Credit loss20.612 Market loss7.716 NII change-0.483 Stress CET15.984 Stress LCR143.877 Judged comparison and execution check. Table 14 and Figure 7 report the judge scores. The pattern is similar to Application 1. RAG-only improves evidence grounding relative to LLM-only, but it does not receive model-output information. The proposed condition receives higher graph and model-grounding scores because the report generator receives the post-run graph context and the executed model output. 13 A′#!+!&% &*+&#,%)' A'+ * C ! A′#!+!&% *%)!& ')! * &" )!+*') * &" '&*!+&,+#&- * &" %.).!# * &" ')! )!+ *') '&*!+ &,+#&-)+ +%.) .!# ,%!% #!(,!!+. !$*-!+ !% *'!!+!&% !$*-!+ !% !$'#$%++!&% &#,% ')! '+ )!+*') '+ +)** +)** Figure 4: Application 2 post-ModelRunGraphRAG paths. The figure displays selected paths from the returned graph result after model execution. Nodes are rendered as circular markers. Table 14: Application 2 judge scores. ConditionEvidenceGraphModelNumericCoherenceAudit LLM-only0.01.00.05.04.02.0 RAG-only5.01.00.05.05.05.0 Model-grounded GraphRAG5.04.05.05.04.05.0 Table 15: Execution checks. ApplicationSelected modelSelected specification Graph clearedOutput points Application 1DSGE-liteDSGE-lite specification 004 YesYes Application 2Regime SwitchingRegime-switching specification 005 YesYes 14 1 1111 "$%&"$*"! 1 1 1 " "'&#'&'!&% ##&"! )'&$ %(&! "#& #$ $&%#$ !'!!%#$ #"%&"'&"($& Figure 5: Application 2 executed regime-switching model path. 000000000 # #('$('("'& '%&& '%&& %' #&& %' #&& " '%&& '%&& $$ '#""&'%&&!'%&%#!!# )('#" Figure 6: Application 2 bank stress metrics from model execution. 15 L& $# !$ ! ! #$"!"&# % ! $! ! $# Figure 7: Application 2 judge scores by condition. 4.3 Discussion The results support three methodological points. First, changing the information set changes what the report can be grounded in. LLM-only can produce a coherent narrative, but it has no retrieved evidence or model execution object. RAG-only generation can cite text evidence, but it is not given a model specification or a model run. Model-grounded GraphRAG adds these objects through the graph and the computational execution layer. Second, the knowledge graph is useful because it separates retrieval from computation. Text spans support narrative statements. Model specifications define what can be executed. Implementation functions compute the output. Model-run and output-point nodes record what was computed. This separation is the main methodological object of the experiment. Third, the judge scores should be read as checks of grounding structure, not as economic performance measures. The numerical values are produced by the experimental model layer. They are not estimates of actual U.S. inflation, Treasury yields, commercial real estate prices, or bank losses. Their role is to make the model-grounding path observable: the selected specification, the executable implementation, the model run, and the output points can be inspected after report generation. The graph figures provide a visual check of this path. In both applications, the post-run graph includes aModelRunand output nodes, and the displayed path connects them to the selected model specification. The figures should therefore be interpreted as diagnostics for the method. They are not independent empirical claims about the two economic settings. 5 Conclusion This study presents a model-grounded RAG-based AI economist for economic report and stress-test narrative generation. The framework separates the planning, evidence retrieval, 16 graph retrieval, model selection, model execution, and report generation processes. In the two applications, RAG-only generation improves textual grounding, whereas model-grounded GraphRAG adds graph and model grounding through selected model specifications and executed model outputs. The central finding is qualitative: requiring the report generator to use retrieved evidence, post-run graph paths, and executed model outputs changes the structure of the generated economic report. It moves the output from a fluent narrative to a traceable economic analysis. References Eason Chen, Chuangji Li, Shizhuo Li, Zimo Xiao, Jionghao Lin, and Kenneth R. Koedinger. Comparing rag and graphrag for page-level retrieval question answering on math textbook, 2025. arXiv: 2509.16780. 3 Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. LightRAG: Simple and fast retrieval-augmented generation. In Findings of the Association for Computational Linguistics (EMNLP). Association for Computational Linguistics, 2025. 3 Bernal Jimenez Gutierrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. HippoRAG: Neurobiologically inspired long-term memory for large language models. In Annual Con- ference on Neural Information Processing Systems (NeurIPS), 2024. 3 Dongwon Jung, Wenxuan Zhou, and Muhao Chen. Code execution as grounded supervision for llm reasoning, 2025. arXiv: 2506.10343. 3 Anton Korinek. Generative ai for economic research: Use cases and implications for economists. Journal of Economic Literature, 61(4):1281–1317, 2023. 2 Anton Korinek. Ai agents for economic research. Technical report, National Bureau of Economic Research, 2025. Working Paper Series. 2 Yassir Lairgi, Ludovic Moncla, Khalid Benabdeslem, R ́emy Cazabet, and Pierre Cl ́eau. ATOM: AdapTive and OptiMized dynamic temporal knowledge graph construction using LLMs. In Findings of the Association for Computational Linguistics: EACL 2026. Association for Computational Linguistics, 2026. 3 Lei Liang, Zhongpu Bo, Zhengke Gui, Zhongshu Zhu, Ling Zhong, Peilong Zhao, Mengshu Sun, Zhiqiang Zhang, Jun Zhou, Wenguang Chen, Wen Zhang, and Huajun Chen. Kag: Boosting llms in professional domains via knowledge augmented generation. In Companion Proceedings of the ACM on Web Conference, 2025. 3 Aditi Singh, Abul Ehtesham, Saket Kumar, Tala Talaei Khoei, and Athanasios V. Vasilakos. Agentic retrieval-augmented generation: A survey on agentic rag, 2026. arXiv: 2501.09136. 3 Yucheng Yang, Yue Pang, Guanhua Huang, and Weinan E. The knowledge graph for macroeconomic analysis with alternative big data, 2020. arXiv: 2010.05172. 3 17 Dalong Zhang, Jun Xu, Jun Zhou, Lei Liang, Lin Yuan, Ling Zhong, Mengshu Sun, Peilong Zhao, QiWei Wang, Xiaorui Wang, Xinkai Du, YangYang Hou, Yu Ao, ZhaoYang Wang, Zhengke Gui, ZhiYing Yi, Zhongpu Bo, Haofen Wang, and Huajun Chen. Kag-thinker: Interactive thinking and deep reasoning in llms via knowledge-augmented generation, 2025. arXiv: 2506.17728. 3 Stephan Zheng, Alexander Trott, Sunil Srinivasa, David C. Parkes, and Richard Socher. The ai economist: Taxation policy design via two-level deep multiagent reinforcement learning. Science Advances, 8(18), 2022. 2 18 A Generated Reports for Application 1 Application 1, LLM-only report Economist Report. The persistence of services inflation in the U.S. economy remains a concern for policymakers at the Federal Reserve. Despite gradual cooling in labor-market conditions, inflationary pressures continue to be felt across several service sectors. From a macroeconomic perspective, the scenario suggests that services inflation may remain elevated even as labor demand cools. The Federal Reserve’s restrictive policy stance is therefore likely to remain in place until the disinflation process becomes clearer. Higher Treasury yields are a direct implication of this policy environment. They raise borrowing costs, slow real activity, and increase the risk that monetary policy becomes too restrictive if labor-market conditions weaken more quickly than expected. The report therefore emphasizes the tension between inflation control and real-activity risk. The Federal Reserve would need to monitor services inflation, labor-market indicators, and financial conditions as the scenario evolves. Application 1, Document-RAG report The scenario describes persistent services inflation, gradual labor-market cooling, restrictive Federal Reserve policy, higher Treasury yields, and slower real activity. The retrieved evidence provides text support for this narrative. EV1571 states that a more restrictive policy path usually lifts nominal yields, although the term premium may vary by regime. EV841 links oil-related shocks to headline inflation before pass-through fades. EV1161 and EV1321 describe credit-loss and commercial real estate channels, while EV1801 and EV721 add broader macro-financial transmission. The Document-RAG report therefore gives an evidence-backed narrative. It does not claim that a model has been executed. Its role is to show how retrieved text changes the report relative to LLM-only generation. Application 1, model-grounded GraphRAG report Application 1 is analyzed using DSGE-lite specificationMSPEC_DSGE_LITE_004. The selected specification maps the inflation and policy scenario to the DSGE-lite Python implementation. The post-ModelRunGraphRAG result connects the scenario to the selected model specification, the implementation function, the model run, and model-output nodes. This graph path is the source of the model-grounded part of the report. The selected model is appropriate because the scenario involves inflation dynamics, monetary-policy reaction, Treasury-yield responses, and real-activity feedback. The model-output summary shows that headline inflation peaks at 0.229 and troughs at -0.455. The federal funds rate peaks at 30.951 and troughs at -5.144. The U.S. ten-year yield peaks 19 at 17.169 and troughs at 2.731. These values are experimental model outputs, not forecasts of actual U.S. data. The economic interpretation is that persistent services inflation can keep the policy stance restrictive, raise nominal yields, and tighten financial conditions. The report separates the qualitative scenario logic from the model-output path and preserves model-run information for traceability. B Generated Reports for Application 2 Application 2, LLM-only report Economist Report: U.S. Commercial Real Estate Refinancing Stress. Higher long-term interest rates and wider credit spreads create refinancing pressure for U.S. commercial real estate borrowers. The increase in borrowing costs can make existing debt harder to roll over and can reduce demand for new lending. Banks may face higher credit losses if property income is insufficient to absorb refinancing costs. Wider credit spreads can also make lenders more cautious, which may further restrict credit availability. Reduced capital and liquidity buffers would weaken the ability of banks to absorb losses. Policymakers and bank risk managers would therefore monitor refinancing volumes, property valuations, funding costs, and liquidity indicators. Application 2, Document-RAG report The scenario of higher long-term rates and wider credit spreads puts pressure on commercial real estate refinancing. It can increase bank credit losses and reduce capital and liquidity buffers. Evidence EV1967 states that funding stress and confidence effects can increase uninsured deposit outflows. EV2028 states that deposit outflows increase net cash outflows and reduce liquidity coverage ratios. EV2107, EV1197, and EV1267 reinforce the deposit-outflow channel. The Document-RAG report therefore grounds the stress narrative in retrieved text. It does not receive the selected model specification or model-output values. Its contribution is textual grounding rather than model grounding. Application 2, model-grounded GraphRAG report Application 2 is analyzed using regime-switching specificationMSPEC_REGIME_SWITCHING_005. The selected specification maps the refinancing-stress scenario to the Regime Switching Python implementation. 20 The post-ModelRunGraphRAG result connects the scenario node, the selected model speci- fication, the implementation function, the model run, output nodes, and bank-metric nodes. The model is appropriate because the scenario involves a transition from normal refinancing conditions to stress conditions in which credit spreads, refinancing costs, and bank-loss channels become nonlinear. The model-output summary shows that the CRE price path peaks at -5.469 and troughs at -16.824. The credit spread peaks at 152.971 and troughs at 71.982. The bank funding spread peaks at 42.760 and troughs at 30.598. The stress module reports PD of 0.049, LGD of 0.467, credit loss of 20.612, market loss of 7.716, NII change of -0.483, stress CET1 of 5.984, and stress LCR of 143.877. The model-grounded interpretation is that higher long-term rates and wider credit spreads reduce refinancing capacity. Lower property valuations and tighter funding conditions increase credit-loss pressure. Liquidity effects appear through deposit-outflow and funding-spread channels. The numerical values are experimental outputs, but the report is traceable to retrieved evidence, graph paths, a selected model specification, and model execution. C Additional Experiment Settings Table 16 summarizes the main settings used in the experiment. These settings are reported to clarify the information environment under which the generated reports and judge scores were obtained. Table 16: Additional experiment settings. SettingValue LLM settingLocal llama3.1:8b model, temperature 0, seed 42. Model horizon12-step synthetic model path. Text evidence returnedSix evidence spans per application. Executable model contextUp to 12 model-specification records before model selection. Model catalog size14 executable families and 30 specifications per family. Graph resetThe graph is cleared before each application is loaded. D Prompt Structure and Judging Criteria Table 17 records the inputs supplied to the report-generation prompts. The purpose is to make the comparison transparent. The prompts differ by information set, not by the general writing objective. 21 Table 17: Prompt inputs used by the three report-generation conditions. Input objectLLM-onlyRAG-onlyModel-grounded GraphRAG Scenario narrativeYesYesYes Retrievedevidence text NoYesYes Pre-run graph pathsNoNoYes Selected model speci- fication NoNoYes Structured execution request NoNoYes Model-run objectNoNoYes Model-output tableNoNoYes Post-run graph pathsNoNoYes Table 18 summarizes the scoring dimensions. The judge receives the generated report and the information set available to the condition. A high model-grounding score requires an executed model object or model-output information. Mentioning a model name without model execution is not sufficient. Table 18: Judge criteria. CriterionMeaning Evidence groundingThe report uses retrieved evidence rather than unsupported claims. Graph reasoningThe report uses graph paths or relationships when they are supplied. Model groundingThe report is linked to a selected model specification and executed output. Numerical disciplineNumerical statements are consistent with the information supplied to the condition. Economic coherenceThe economic interpretation is internally consistent. AuditThe report exposes enough information to trace the source of its claims. E Model Execution Records Table 19 reports the structured model-selection outputs used by the proposed condition. The selected model specification is a graph object. The implementation family is the executable family used for model execution. Table 20 summarizes the graph write-back step. The proposed condition uses the graph 22 Table 19: Model-selection and execution records. ApplicationSelected modelSelected specificationImplementation family Main outputs Application 1DSGE-liteDSGE-lite specification 004 DSGE-liteMacro path. Application 2Regime SwitchingRegime-switching spec- ification 005 Regime Switch- ing Stress path and bank metrics. after this write-back, so the report generator can see the selected model and the computed output objects. Table 20: Graph write-back objects created after model execution. ObjectRole ModelRunRecords that a selected model specification was executed for the application. ForecastOutcomeStores the summary of the model-output path. ModelOutputPointStores horizon-level values and links them to output variables. BankMetricStores stress metrics used in the stress-test narrative. Table 21 reports the execution trace used in both applications. The trace records the order of the agent calls and the graph operations. It is included to clarify that the proposed condition uses graph write-back before final report generation. Table 21: Agent execution trace. StepComponentOutputStatus Graph resetGraph store commandExisting graph contents removed before loading the application. Completed Graph loadingGraph loaderEvidence rows, temporal facts, model catalog, scenario node, and scenario-shock relationships loaded. Completed PlanningLLM plannerAnalysis objective, target variables, and model-selection objective returned. Completed Text retrievalEvidence querySix evidence spans returned to the RAG-only and proposed conditions. Completed Graph retrieval before execu- tion Graph path queryPre-run graph paths and model context returned. Completed Model selectionLLM model selectorStructured execution request returned.Completed Model executionModel execution layerModel-output table and stress metrics computed. Completed Graph write-backGraph writerModel-run and output nodes recorded in the graph. Completed Graph retrieval after executionGraph path queryPost-run graph paths returned.Completed Report generation and judgingLLM report generator and judge Three reports and judge scores returned. Completed 23