ProtoReasoning: General-purpose Reasoning Skills Honed Through Logic And Planning
3 main points
✔️ We propose a "ProtoReasoning" framework that leverages prototypes to enable large-scale language models to effectively learn reasoning capabilities in different problem domains
✔️ We develop a prototype representation of logical reasoning and planning using Prolog and PDDL, which allows models toEnhanced reasoning capabilities for structurally different problems
✔️ Experimental results confirm that the proposed framework has better generalization capabilities than natural language representations and improved performance for different tasks
ProtoReasoning: Prototypes as the Foundation for Generalizable Reasoning in LLMs
written by Feng He, Zijun Chen, Xinnian Liang, Tingting Ma, Yunqi Qiu, Shuangzhi Wu, Junchi Yan
(Submitted on 18 Jun 2025)
Comments: Published on arxiv.
Subjects: Computation and Language (cs.CL)
code：
The images used in this article are from the paper, the introductory slides, or were created based on them.
Summary
This paper proposes a new framework, ProtoReasoning, for giant language models (LLMs) to have the ability to generalize and solve different problems ProtoReasoning leverages prototypical representations to mimic human cognition and problem solving abilities It aims to At the heart of this framework is a prototype generation and verification system using formal languages such as Prolog and PDDL.
Prolog specializes in logical reasoning and solves problems by decomposing them into logical facts and rules. PDDL, on the other hand, provides a representation for planning and models the process similar to human planning thinking. This allows for training models with flexible reasoning capabilities for different problem domains.
Experimental results show that this approach performs better than traditional methods and can effectively improve logical reasoning and planning abilities. This research provides a new foundation for improving the adaptive capacity of LLMs and is expected to have broader applications.
Proposed Methodology
A new framework called "ProtoReasoning" is proposed to improve LLMs' capabilities. This framework aims to mimic human problem-solving capabilities using prototype representations. Specifically, we utilize the formal languages Prolog and PDDL to create a system for generating prototypes and verifying their correctness.
Prolog is mainly dedicated to logical reasoning and has a method of solving problems by breaking them down into logical facts and rules. PDDL, on the other hand, provides a representation for planning and models planning as humans would do it.
By combining these two languages, it is possible to give models the ability to reason flexibly in different problem domains.
Experiments
The experiments in this study test the effectiveness of the proposed method, ProtoReasoning, in improving reasoning ability.
First, natural language problems were converted into logical code, using the format Prolog for logical reasoning and PDDL for planning tasks. This allowed us to accurately understand the structure of the problem and build a data set that could automatically generate the correct answer. Finally, approximately 4,200 Prolog questions and 2,400 PDDL questions were prepared and the model was trained in three stages.
After training, we evaluated the performance not only in logical reasoning and planning, but also in general knowledge and math benchmarks. The results showed a significant performance improvement of 4.7% for logical reasoning and 6.3% for the planning task. Ablation experiments also showed that learning with logical representations is at least as effective as learning with natural language alone.
The results of this experiment strongly support the effectiveness of learning logical structures.
Conclusion
In this paper, we proposed a new framework called "ProtoReasoning" that aims to improve the ability of large language models (LLMs) to generalize and solve various problems.
ProtoReasoning utilizes prototypical representations and introduces methods that mimic human cognition and problem-solving abilities. At the heart of the framework is a prototype generation and validation system using Prolog and PDDL, a formal language.
Prolog specializes in logical reasoning and solves problems by decomposing them into logical facts and rules; PDDL provides a representation for planning and models the process similar to human planning thinking. Combining these languages allows for flexible reasoning capabilities in different problem domains.
Experimental results show that this approach performs better than traditional methods and can effectively improve logical reasoning and planning abilities. This research provides a new foundation for improving the adaptive capacity of LLMs and is expected to have broader applications.
Categories related to this article
