ProRAC: A Neuro-symbolic Method for Reasoning about Actions with LLM-based Progression and Search

ACL ARR 2026 January Submission7103 Authors

Keywords: neuro-symbolic, Reasoning about Actions, Planning, Large Language Models, Prompting, Finetuing

Abstract:

Reasoning about actions and change (RAC) plays an important role in AI. It involves reasoning about preconditions and effects of actions, and has applications in planning.
Despite recent advances of large language models (LLMs) in natural language processing, evaluations on several RAC benchmarks demonstrate that LLMs face significant challenges in RAC. Besides, effective methods for improving their RAC capabilities, especially neuro-symbolic approaches, remain largely unexplored.
In this paper, we propose **ProRAC** (**Pro**gression-based **R**easoning about **A**ctions and **C**hange), a neuro‐symbolic framework that leverages LLMs to tackle RAC and planning problems through an agentic, modular design.
Central to ProRAC is a unified state progression framework that operates in two modes: a reasoning mode that sequentially executes actions to validate states and answer queries, and a searching mode that integrates this progression into an A* search guided by a fine-tuned heuristic model to find plans. Extensive evaluations across multiple RAC benchmarks and planning domains demonstrate that ProRAC significantly outperforms existing methods, exhibiting robust performance across different model backbones and task types.

Paper Type: Long

Research Area: Mathematical, Symbolic, Neurosymbolic, and Logical Reasoning

Research Area Keywords: neurosymbolic reasoning, symbolic AI

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 7103