RSAT: Structured Attribution Makes Small Language Models Faithful Table Reasoners
Tiny models now cite table cells with 99.2% accuracy, outperforming post-hoc methods 3.7x
When a language model answers a table-based question, users currently cannot verify which cells informed each reasoning step. A new paper from researchers Jugal Gajjar and Kamalasankari Subramaniakuppusamy introduces RSAT (Reasoning with Structured Attribution for Tables), a method that trains small language models (SLMs) to produce step-by-step reasoning with explicit, cell-level citations grounded in the table evidence. The approach uses a two-phase pipeline: Phase 1 (Supervised Fine-Tuning) teaches models to output a structured JSON format from verified reasoning traces, and Phase 2 (GRPO optimization) applies a composite reward centered on NLI-based faithfulness, citation validity, and citation parsimony.
Across six models from two families—Qwen 2.5 (1.5B/3B/7B) and Llama 3 (1B/3B/8B)—RSAT improved faithfulness by 3.7× over SFT alone (from 0.224 to 0.826) and achieved near-perfect citation validity of 0.992. Critically, post-hoc attribution methods (adding citations after generation) collapsed below 13% format success, confirming that attribution must be integrated into the reasoning process, not retrofitted. Ablation studies showed the faithfulness reward is essential: removing it dropped faithfulness from 0.97 to 0.03. The paper was accepted at the SURGeLLM Workshop at ACL 2026.
- RSAT trains 1-8B models (Qwen 2.5 & Llama 3) to output cell-level citations per reasoning step
- Improves faithfulness 3.7× over SFT alone (0.224 → 0.826) with near-perfect citation validity (0.992)
- Post-hoc attribution fails (<13% success), proving citations must be built into reasoning, not added after
Why It Matters
Enables small, efficient models to produce verifiable table reasoning—critical for trust in enterprise analytics and automated data QA.