RSAT: Structured Attribution Makes Small Language Models Faithful Table Reasoners

TL;DR

RSAT is a training method for small language models that enhances their ability to produce faithful, step-by-step table reasoning with cell-level citations, improving transparency and trustworthiness.

Contribution

The paper introduces RSAT, a novel training approach that integrates structured reasoning and faithfulness rewards into small language models for better table question answering.

Findings

RSAT improves faithfulness 3.7× over standard fine-tuning.

Near-perfect citation validity of 0.992 achieved.

Post-hoc attribution methods are less effective than integrated training.

Abstract

When a language model answers a table question, users have no way to verify which cells informed which reasoning steps. We introduce RSAT, a method that trains small language models (SLMs, 1-8B) to produce step-by-step reasoning with cell-level citations grounded in table evidence. Phase 1 (SFT) teaches a structured JSON output format from verified reasoning traces. Phase 2 (GRPO) optimizes a composite reward centered on NLI-based faithfulness, alongside citation validity and parsimony. Across six models from two families-Qwen 2.5 (1.5B/3B/7B) and Llama 3 (1B/3B/8B)-RSAT improves faithfulness 3.7$\times$ over SFT alone (0.224$\rightarrow$0.826), with near-perfect citation validity (0.992). Post-hoc attribution collapses below 13% format success, confirming that attribution must be integrated into reasoning, not retrofitted. Ablations show the faithfulness reward is essential: removing it drops faithfulness from 0.97 to 0.03.