RSAT: Small Language Models Gain Reliable Table Reasoning via Structured Attribution
The RSAT method enhances the ability of small language models to perform reliable table-based reasoning by integrating cell-level citations, thereby increasing trustworthiness. This addresses a clinical issue where users previously could not verify the source of a model's reasoning.
Vad har hänt
Researchers have introduced RSAT (Reasoning with Structured Attribution), a method that trains small language models (SLM, 1-8B parameters) to generate step-by-step reasoning based on tabular data. RSAT produces output in a structured JSON format that includes specific cell-level citations from the table. The method was developed in two phases: first, supervised fine-tuning (SFT) with verified reasoning traces, followed by GRPO (Group Relative Policy Optimisation) using a composite reward based on NLI (Natural Language Inference) faithfulness, citation validity, and conciseness.
Key facts
| Metod | RSAT |
|---|---|
| Modellstorlek | 1-8 miljarder parametrar (Qwen 2.5, Llama 3) |
| Trovärdighetsförbättring | 3.7x över SFT |
| Citatgiltighet | 0.992 |
| Post-hoc attribuering formatframgång | <13% |
”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.”
”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).”
Varför det spelar roll
The problem RSAT addresses is that language models traditionally cannot disclose which specific data points they used in their reasoning steps, making it difficult for users to verify correctness. By enforcing cell-level citations, the reasoning becomes transparent and auditable. This is critical for applications requiring high reliability, such as finance, law, and medicine, where incorrect answers can have serious consequences. The method demonstrates that attribution must be an integrated part of the model's reasoning rather than a post-hoc analysis.
Vem påverkas
RSAT primarily concerns developers and researchers in Natural Language Processing (NLP) working with small language models and tabular data. Companies implementing AI solutions where data transparency and verifiability are central will also be affected. Users of AI systems who need to understand how a model reaches its conclusions will benefit from increased reliability and traceability.
EU-status
Ej relevant för EU-status.
Mer att veta
Post-hoc attribution, where citations are added after the fact, has proven ineffective, with under 13% success in formatting output correctly. Ablation tests confirmed that the faithfulness reward is critical; without it, faithfulness dropped from 0.97 to 0.03.
Quick answers about this story
Vad har hänt?
När hände det?
Varför spelar det roll?
Vilka modeller berörs?
Länken öppnar i nytt fönster och leder till utgivarens egen sida.
Källan har spårats automatiskt från utgivaren via Aheadlines signalkedja.