AI, LLMs, and Decision-Making
A language model writes fluently. Fluency is not evidence. This topic covers ways of tying an analysis to the documents it retrieved and to model computation that can be inspected, so that what the model writes leads to a decision someone can check afterwards. The central result is AI Economist Agent (Kato, 2026, arXiv:2606.20041), an agent framework that generates economic analysis.
Problems addressed
Take economic analysis. An analyst is expected to back a claim with theory and with data. Ask a language model to write the analysis directly and that backing disappears, because nobody can go back and see which source and which calculation produced the conclusion. So I split the work into separate components: planning the analysis, retrieving the evidence, choosing a model, computing, and writing the report. Every numerical claim comes out of an explicit model computation rather than out of the language model. Retrieval uses RAG and a knowledge graph, and each claim in the report is linked back to the documents retrieved and the results computed.
Decision-making itself belongs here too. Gathering similar past cases for each action by vector search and estimating the outcome from them can be formulated as nearest neighbor matching in causal inference (RAG-based policy learning). A separate line of work draws candidate causal relations out of what a language model already knows, in a form a person can inspect.
Core research
AI Economist Agent combines RAG, a knowledge graph, economic models, and LLM agents to generate economic scenario analysis. I evaluated it on two tasks: writing a report on inflation persistence and monetary policy in the United States, and producing the narrative for a bank stress test covering refinancing stress in commercial real estate.
Research thread
- 2026
- Vector Search As Nearest Neighbor Matching: RAG-based Policy Learning in Causal Inference. Formulates RAG-based action selection in the language of causal inference.
- 2026
- Causality Elicitation from Large Language Models. Draws candidate causal relations out of a language model and organizes them so they can be inspected.
- 2026
- Which Algorithm Specification Formats Help Language Models Implement Machine Learning Algorithms?. An experiment measuring how the way an algorithm is written affects the accuracy of the implementations a language model produces. The AI-Ready Knowledge Infrastructure page covers it in detail.
Relation to earlier work
RAG, which backs generation with retrieved documents, was proposed by Lewis and coauthors in 2020 (Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks). Using a knowledge graph to carry the structure of a whole document collection into a summary is the idea behind Graph RAG, from Edge and coauthors (2024, From Local to Global). Both target the retrieval and summarization of documents. My work adds explicit computation from an economic model on top of retrieval, and it keeps numerical claims out of the language model's hands.
Where it applies
The situation this work addresses is one where you would like AI to do the analysis but cannot use the result unless the grounds for the conclusion can be explained. Automating market analysis and report writing falls under this, along with analytical support that draws on internal data alongside a mathematical model, and the design of a review process for what an AI system outputs. For inquiries of this kind, write to mkato-csecon@g.ecc.u-tokyo.ac.jp.