RSAT: Structured Attribution Makes Small Language Models Faithful Table Reasoners

Researchers introduce RSAT, a method that trains small language models (1-8B parameters) to answer table-based questions with step-by-step reasoning and cell-level citations, achieving 3.7x improvement in faithfulness over baseline approaches. The technique uses structured JSON outputs and reinforcement learning to ensure AI reasoning is verifiable and grounded in source data.

RSAT addresses a critical challenge in deploying language models for information retrieval: the inability to verify which data sources informed specific reasoning steps. This transparency gap undermines trust in AI systems, particularly in domains like financial analysis, legal research, and scientific investigation where source attribution directly impacts decision-making reliability. The research demonstrates that achieving interpretable reasoning requires integrating attribution into the model's reasoning process rather than applying it post-hoc.

The method's two-phase approach reflects evolving best practices in language model fine-tuning. Supervised fine-tuning establishes the structured output format, while group relative policy optimization (GRPO) optimizes for multiple objectives simultaneously: faithfulness measured through Natural Language Inference, citation validity, and output parsimony. The dramatic performance improvements—from 0.224 to 0.826 faithfulness across multiple model scales—suggest the approach generalizes effectively.

For developers and organizations deploying SLMs in production, this work offers practical techniques for building verifiable AI systems without requiring large models. The finding that post-hoc attribution methods collapse to 13% format success proves that trustworthiness cannot be grafted onto reasoning as an afterthought. This has implications for regulatory compliance and user trust, particularly as enterprises increasingly rely on AI for knowledge work.

The research emphasizes that faithful reasoning requires deliberate architectural choices during training. Future work likely extends these methods to other domains beyond table reasoning—document retrieval, code generation, and multi-hop reasoning scenarios where source attribution provides significant value.

• →RSAT improves faithful reasoning in small language models 3.7x over baseline by integrating attribution during training rather than retrofitting it afterward
• →Post-hoc attribution methods achieve only 13% format success, proving that verifiable reasoning must be built into model architecture from the start
• →The method demonstrates consistent improvements across six different SLM configurations (Qwen and Llama families), suggesting broad applicability
• →Faithfulness rewards prove essential: removing them drops performance from 0.97 to 0.03, indicating composite reward optimization is critical for trustworthy outputs
• →Citation validity reaches near-perfect levels (0.992), enabling users to verify which data sources directly informed each reasoning step