让小模型回答表格问题时能精准引用依据，提升可信度
RSAT: Structured Attribution Makes Small Language Models Faithful Table Reasoners
- 训练小模型生成带单元格引用的分步推理
- 引用准确率接近完美（99.2%），可信度提升3.7倍
- 适合需要可解释性表格问答的场景
当语言模型回答表格问题时，用户无法判断哪些单元格影响了推理步骤。我们提出RSAT方法，训练小型语言模型（1-8B）生成带有单元格级引用的分步推理。第一阶段（SFT）通过已验证的推理轨迹学习结构化JSON输出格式；第二阶段（GRPO）优化以NLI为基础的可信度奖励，同时兼顾引用有效性和简洁性。在两个系列共六种模型（Qwen 2.5: 1.5B/3B/7B；Llama 3: 1B/3B/8B）上，相比仅用SFT，RSAT使可信度从0.224提升至0.826，引用准确率达0.992。事后归因成功率低于13%，表明归因必须嵌入推理过程而非事后添加。消融实验显示，移除可信度奖励会使可信度从0.97骤降至0.03。
原文摘要 · Abstract (English)
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.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。