RSAT: Structured Attribution Makes Small Language Models Faithful Table Reasoners

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

RSAT is a novel method designed to enhance the faithfulness of Small Language Models (SLMs, 1–8B) when answering table-based questions by integrating structured, cell-level attribution into their reasoning processes. This two-phase approach first employs Supervised Fine-Tuning (SFT) to teach SLMs a structured JSON output format for reasoning traces. Subsequently, a Grouped Reinforcement Learning with Policy Optimization (GRPO) phase optimizes a composite reward, prioritizing NLI-based faithfulness, citation validity, and parsimony. Evaluated across six models from the Qwen2.5 (1.5B/3B/7B) and Llama3 (1B/3B/8B) families, RSAT significantly improved faithfulness by 3.7×, from 0.224 to 0.826, and achieved a high citation validity of 0.992. The research also demonstrated that post-hoc attribution is ineffective, yielding below 13% format success, underscoring the necessity of integrating attribution directly into the reasoning process. Ablation studies confirmed the critical role of the faithfulness reward, as its removal caused a drastic drop in faithfulness from 0.97 to 0.03.

Key takeaway

For Machine Learning Engineers developing explainable AI for tabular data, you should prioritize integrating attribution directly into your Small Language Models' (SLMs) training pipeline. Retrofitting attribution post-training is largely ineffective, as shown by its low format success. Instead, adopt a multi-phase training approach like RSAT, incorporating structured output formats and NLI-based faithfulness rewards to achieve verifiable, step-by-step reasoning with high citation validity. This approach ensures your models provide transparent and trustworthy answers.

Key insights

Integrating structured, cell-level attribution directly into SLM reasoning significantly boosts faithfulness and verifiability.

Principles

• Attribution must be integrated, not retrofitted, into reasoning.
• NLI-based faithfulness rewards are essential for verifiable reasoning.
• Structured output formats improve model interpretability and auditability.

Method

RSAT trains SLMs in two phases: SFT for structured JSON output from verified traces, then GRPO with a composite reward for NLI-based faithfulness, citation validity, and parsimony.

In practice

• Train SLMs to produce structured JSON outputs for reasoning.
• Incorporate NLI-based faithfulness rewards during optimization.
• Prioritize integrated attribution over post-hoc methods for transparency.

Topics

• Small Language Models
• Table Reasoning
• Attribution
• Faithfulness
• Reinforcement Learning
• Structured Output

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Faithfulness Serum: Mitigating the Faithfulness Gap in Textual Explanations of LLM Decisions via Attribution Guidance — LLM-generated explanations often lack epistemic faithfulness, but a new method improves it via attention-level guidance.
• SLMJury: Can Small Language Models Judge as Well as Large Ones? — Small Language Models can effectively judge model outputs, but their performance is highly dependent on the task domain and judging paradigm.
• Towards On-Policy SFT: Distribution Discriminant Theory and its Applications in LLM Training — On-Policy SFT, enabled by Distribution Discriminant Theory, matches RL generalization with SFT efficiency.
• LoRi: Low-Rank Distillation for Implicit Reasoning — Hidden-state reasoning trajectories exhibit low-rank structure, enabling efficient distillation for implicit reasoning.
• Pruning via Causal Attribution Preserves Reasoning Performance in Large Language Models — Causal Attribution Pruning (CAP) uses interventional head masking to identify and protect critical attention heads for reasoning performance.
• LaCy: What Small Language Models Can and Should Learn is Not Just a Question of Loss — SLMs can improve factual accuracy by learning to delegate specific tokens rather than predicting all tokens.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.