Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neuro-Symbolic Integration Brings Causal and Reliable Reasoning Proofs

Sen Yang, Xin Li, Leyang Cui, Lidong Bing, Wai Lam

TL;DR

The paper addresses the challenge of achieving accurate and interpretable reasoning with large language models by separating problem translation from reasoning execution. It introduces CaRing, a neuro-symbolic framework that translates problems into Prolog representations (SymGen) and uses a Prolog-based inference engine (SymInfer) to produce causal, deterministic proofs via a reasoning tracer. Across ProofWriter, PrOntoQA, and GSM8K, CaRing yields state-of-the-art or near state-of-the-art results in both final answers and the quality of reasoning proofs, demonstrating that solver-generated logs can be transformed into human-readable and verifiable explanations. This approach offers a scalable path to transparent reasoning without sacrificing performance, though it relies on strong symbolic representations and LLMs for initial problem translation.

Abstract

Two lines of approaches are adopted for complex reasoning with LLMs. One line of work prompts LLMs with various reasoning structures, while the structural outputs can be naturally regarded as intermediate reasoning steps. Another line of work adopt LLM-free declarative solvers to do the reasoning task, rendering higher reasoning accuracy but lacking interpretability due to the black-box nature of the solvers. Aiming to resolve the trade-off between answer accuracy and interpretability, we present a simple extension to the latter line of work. Specifically, we showcase that the intermediate search logs generated by Prolog interpreters can be accessed and interpreted into human-readable reasoning proofs. As long as LLMs correctly translate problem descriptions into Prolog representations, the corresponding reasoning proofs are ensured to be causal and reliable. On two logical reasoning and one arithmetic reasoning datasets, our framework obtains significant improvements in terms of both answer accuracy and reasoning proof accuracy. Our code is released at https://github.com/DAMO-NLP-SG/CaRing

Neuro-Symbolic Integration Brings Causal and Reliable Reasoning Proofs

TL;DR

The paper addresses the challenge of achieving accurate and interpretable reasoning with large language models by separating problem translation from reasoning execution. It introduces CaRing, a neuro-symbolic framework that translates problems into Prolog representations (SymGen) and uses a Prolog-based inference engine (SymInfer) to produce causal, deterministic proofs via a reasoning tracer. Across ProofWriter, PrOntoQA, and GSM8K, CaRing yields state-of-the-art or near state-of-the-art results in both final answers and the quality of reasoning proofs, demonstrating that solver-generated logs can be transformed into human-readable and verifiable explanations. This approach offers a scalable path to transparent reasoning without sacrificing performance, though it relies on strong symbolic representations and LLMs for initial problem translation.

Abstract

Two lines of approaches are adopted for complex reasoning with LLMs. One line of work prompts LLMs with various reasoning structures, while the structural outputs can be naturally regarded as intermediate reasoning steps. Another line of work adopt LLM-free declarative solvers to do the reasoning task, rendering higher reasoning accuracy but lacking interpretability due to the black-box nature of the solvers. Aiming to resolve the trade-off between answer accuracy and interpretability, we present a simple extension to the latter line of work. Specifically, we showcase that the intermediate search logs generated by Prolog interpreters can be accessed and interpreted into human-readable reasoning proofs. As long as LLMs correctly translate problem descriptions into Prolog representations, the corresponding reasoning proofs are ensured to be causal and reliable. On two logical reasoning and one arithmetic reasoning datasets, our framework obtains significant improvements in terms of both answer accuracy and reasoning proof accuracy. Our code is released at https://github.com/DAMO-NLP-SG/CaRing
Paper Structure (35 sections, 1 equation, 4 figures, 6 tables)

This paper contains 35 sections, 1 equation, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Explainable Complex Reasoning
  4. Neuro-symbolic Reasoning
  5. Method
  6. SymGen: Symbolic Representation Generator
  7. SymInfer: Symbolic Inference Engine
  8. Reasoning Tracer
  9. Search Strategy
  10. Experiments
  11. Experimental Settings
  12. Implementation
  13. SymGen
  14. SymInfer
  15. Datasets
  16. ...and 20 more sections

Figures (4)

  • Figure 1: Two examples of complex/structured reasoning problems from ProofWriter and GSM8K, respectively. The reasoning proofs in such problems formulate directed acyclic graphs (DAGs) in a multi-step and multi-premise manner.
  • Figure 2: Illustration of our CaRing framework, consisting of a Logic component and a Control component.
  • Figure 3: Illustrations of how causality and reliability play important roles in reasoning. LLMs may be (i) non-causal by selecting redundant premises or ignoring relevant ones and (ii) non-reliable by hallucinating erroneous contents.
  • Figure 4: Answer accuracy when reasoning problems become more complex.