ReSS: Symbolic Scaffolds for Tabular Reasoning

🔑 Enhanced Key Takeaways

• •ReSS addresses the 'black-box' reasoning problem in LLMs by enforcing a neuro-symbolic bridge, where the decision tree acts as a hard constraint on the LLM's latent reasoning space.
• •The framework utilizes a novel 'scaffold-invariant' data augmentation technique that synthetically perturbs tabular data while preserving the underlying decision logic, significantly improving model robustness against adversarial table inputs.
• •The introduced faithfulness metrics specifically quantify the alignment between the LLM's generated natural language explanation and the ground-truth decision path, effectively penalizing 'hallucinated' reasoning steps that deviate from the symbolic scaffold.

🛠️ Technical Deep Dive

• •Architecture: Employs a two-stage pipeline: (1) Symbolic Extraction Module that converts tabular data into a decision tree representation; (2) Scaffold-Guided Decoder that uses the tree path as a prefix constraint during LLM inference.
• •Scaffold-Invariant Augmentation: Implements a perturbation strategy that swaps non-critical feature values while maintaining the decision tree's leaf node outcome, forcing the LLM to focus on causal features.
• •Faithfulness Metrics: Defines 'Explanation Necessity' as the probability that the conclusion changes if the explanation is removed, and 'Explanation Sufficiency' as the probability that the explanation alone predicts the correct label.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline