ReSS Framework: A New Approach to Tabular Data Reasoning Combining Symbolic Scaffolds and Neural Networks

The ReSS (Reasoning via Symbolic Scaffold) framework uses decision trees to extract symbolic scaffolds that guide Large Language Models (LLMs) in generating faithful reasoning, addressing the dual challenges of accuracy and interpretability in tabular data prediction. In medical and financial benchmark tests, it achieves an accuracy improvement of up to 10% compared to traditional methods while ensuring reasoning consistency and interpretability.

Tabular data dominates high-risk domains such as healthcare and finance, where prediction models are required to have both high accuracy and human-understandable faithful reasoning. Traditional symbolic models (e.g., decision trees) provide verifiable logic but have limited semantic expression capabilities; general-purpose LLMs have strong language generation capabilities but require extensive domain-specific fine-tuning. Additionally, scaling the construction of high-quality training data and ensuring reasoning consistency are two major challenges.

The ReSS framework bridges the gap between symbolic models and neural networks through a three-step process:

1. Symbolic Scaffold Extraction: Use decision trees to extract instance-level decision paths from training data as structured guidance;
2. Faithful Reasoning Generation: Input the scaffold, input features, and labels into the LLM to guide the generation of natural language reasoning that follows decision logic, eliminating hallucinations;
3. Specialized Fine-Tuning and Enhancement: Fine-tune pre-trained LLMs using the generated high-quality dataset, and expand training data through scaffold-invariant data augmentation to improve generalization capabilities.

In multiple benchmark tests in the healthcare and finance domains, the ReSS model performed exceptionally well: compared to traditional decision trees and standard fine-tuning methods, it achieved up to 10% higher accuracy without sacrificing interpretability—generated reasoning is more faithful and consistent. Experiments show that neural networks guided by symbolic scaffolds can retain both the expressive power of neural networks and the interpretability advantages of symbolic models.

The core innovations of ReSS include:

• Faithfulness Evaluation Metrics: Proposed three quantitative metrics—hallucination rate (measuring the proportion of factual inconsistencies), explanation necessity (eliminating redundancy), and explanation sufficiency (covering key information)—to build a comprehensive credibility evaluation framework;
• Knowledge Distillation Paradigm: Transfer the interpretability of symbolic models to neural networks while retaining the latter's expressive power;
• Scaffold-Invariant Data Augmentation: Expand data through semantic-preserving decision path transformations to alleviate the problem of scarce annotations.

ReSS has significant value for industries requiring highly interpretable predictions:

• Medical Diagnosis: Provides clear and verifiable reasoning chains, enhancing doctors' trust in AI systems;
• Financial Risk Control: Generates faithful reasoning to meet regulatory compliance requirements while maintaining high prediction accuracy;
• Insurance Claims and Credit Approval: Balances automation efficiency and decision transparency.

The ReSS framework paves a new path for tabular data prediction tasks that balances accuracy, interpretability, and scalability. By intelligently using symbolic scaffolds, it successfully combines the interpretability of symbolic models with the strong expressive power of neural networks, providing a more reliable technical solution for AI applications in high-risk domains.