Interpretable Decision-Making with Large Language Models

The field of artificial intelligence is moving towards more interpretable and transparent decision-making systems. Recent developments have focused on leveraging large language models (LLMs) to induce decision trees, build thematic trees, and create agentic classification trees. These approaches aim to provide human-readable reasoning traces, explicit logic, and transparent decision paths, making it possible to check for biases and data leaks. The use of LLMs is also being explored in data discovery, where they can enhance the search process by allowing researchers to ask questions in natural language. Furthermore, LLMs are being applied to contextual Markov decision processes to compress contextual inputs into low-dimensional, semantically rich summaries, improving decision-making in context-rich environments. Noteworthy papers include:

  • Talking Trees, which explores the use of reasoning-capable LLMs to induce decision trees for small tabular datasets.
  • Question-Driven Analysis and Synthesis, which introduces a novel framework for building interpretable thematic trees with LLMs.
  • ACT, which extends decision-tree methodology to unstructured inputs by formulating each split as a natural-language question.
  • Learning to Decide with Just Enough, which proposes an information-theoretic summarization approach using LLMs to compress contextual inputs.

Sources

Talking Trees: Reasoning-Assisted Induction of Decision Trees for Tabular Data

Question-Driven Analysis and Synthesis: Building Interpretable Thematic Trees with LLMs for Text Clustering and Controllable Generation

First Workshop on Building Innovative Research Systems for Digital Libraries (BIRDS 2025)

ACT: Agentic Classification Tree

From keywords to semantics: Perceptions of large language models in data discovery

Learning to Decide with Just Enough: Information-Theoretic Context Summarization for CDMPs

Pre-Hoc Predictions in AutoML: Leveraging LLMs to Enhance Model Selection and Benchmarking for Tabular datasets

Built with on top of