RSAT: Structured Attribution Makes Small Language Models Faithful Table Reasoners
Jugal Gajjar, Kamalasankari Subramaniakuppusamy
TLDR
RSAT trains small language models to provide faithful, step-by-step table reasoning with verifiable cell-level citations, significantly improving accuracy.
Key contributions
- RSAT trains SLMs (1-8B) for faithful table reasoning with cell-level citations.
- Employs a two-phase method: SFT for structured JSON output and GRPO for reward-based optimization.
- Achieves a 3.7x faithfulness improvement and 99.2% citation validity over SFT alone.
- Demonstrates that integrated attribution is crucial, outperforming post-hoc methods.
Why it matters
This paper addresses a critical transparency issue in LLM table reasoning by enabling verifiable, step-by-step explanations. It significantly enhances the trustworthiness and interpretability of small language models for complex data tasks, making them more reliable for real-world applications.
Original 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.
📬 Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.