Publications

Graph of States: Solving Abductive Tasks with Large Language Models

Yu Luo1, Rongchen Gao1, Lu Teng2, Xidao Wen3, Jiamin Jiang1, Qingliang Zhang1, Yongqian Sun1,*, Shenglin Zhang1, Jiasong Feng4, Tong Liu4, Wenjie Zhang4, Dan Pei5

1Nankai University   2Wenzhou Medical University   3Alibaba Cloud   4Lenovo   5Tsinghua University

ICML 2026

43rd International Conference on Machine Learning, Seoul, South Korea

* Corresponding author

Paper Code 机器之心 Slide Story
Graph of States teaser figure
Graph of States turns abductive reasoning into a directed search over structured belief states, causal relations, and state transitions.

Abstract

Graph of States is a general-purpose neuro-symbolic framework for abductive reasoning with large language models. It grounds multi-agent collaboration in explicit belief states, uses a causal graph and state machine to constrain reasoning transitions, and converts open-ended exploration into a directed search over possible explanations. The framework is designed for complex tasks where agents must collect evidence, revise hypotheses, and identify the most plausible hidden cause behind observed symptoms.

Project Overview

Belief States

The reasoning process is represented as structured states, making collaboration more compact and traceable than long conversational histories.

Causal Transitions

A causal graph and state machine constrain how hypotheses are expanded, refined, and rejected during abductive search.

General Abduction

The design applies beyond microservice diagnosis, supporting broader abductive tasks where observed evidence must be explained by hidden causes.

Method

Graph of States method overview
The overview shows how central and expert agents update the graph-structured belief state through planning, investigation, and state conversion.

Citation

@inproceedings{luo2026graphstates,
  title={Graph of States: Solving Abductive Tasks with Large Language Models},
  author={Luo, Yu and Gao, Rongchen and Teng, Lu and Wen, Xidao and Jiang, Jiamin and Zhang, Qingliang and Sun, Yongqian and Zhang, Shenglin and Feng, Jiasong and Liu, Tong and Zhang, Wenjie and Pei, Dan},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning},
  year={2026}
}