逐步训练用于可解释的引用导向对话：在英语-印地语LLM中将幻觉减少到零
Progressive Training for Explainable Citation-Grounded Dialogue: Reducing Hallucination to Zero in English-Hindi LLMs
AI总结：
本文提出XKD-Dial，一种逐步四阶段训练流程，用于生成可解释的引用导向对话，通过跨语言适应、英语对话SFT与引用定位、双语对话SFT及GRPO对齐与引用感知奖励，减少幻觉并提升印地语能力。
摘要：
知识导向对话系统旨在通过外部知识源生成信息丰富且上下文相关的响应。然而，现有方法大多仅专注于英语，缺乏验证事实性声明的显式引用机制，并且对模型决策过程的透明度有限。我们提出了XKD-Dial，一种逐步四阶段训练流程，用于在双语（英语-印地语）环境中生成可解释的知识导向对话，包括：（1）多语言适应，（2）英语对话SFT与引用定位，（3）双语对话SFT，以及（4）GRPO对齐与引用感知奖励。我们评估了六个模型，涵盖编码器-解码器（250M-3B）和解码器-only（1B-7B）架构，在每个训练阶段。我们的主要贡献是：（i）三种后验可解释性分析——跨注意力对齐、集成梯度归因和基于遮蔽的因果定位——系统性地应用于训练轨迹，揭示引用行为如何学习，而不仅仅是是否学习；（ii）引用导向的SFT在编码器-解码器模型从第2阶段起将幻觉减少到0.0%；（iii）渐进式流程在防止灾难性遗忘的同时提高印地语能力；（iv）较小的模型在SFT后在英语上与较大的模型相当；（v）GRPO在结构化引用任务中对well-designed SFT提供边际改进。我们评估了六个自动指标（BLEU、ROUGE、BERTScore、FactScore、Citation-F1和幻觉率）。
英文摘要：
Knowledge-grounded dialogue systems aim to generate informative, contextually relevant responses by conditioning on external knowledge sources. However, most existing approaches focus exclusively on English, lack explicit citation mechanisms for verifying factual claims, and offer limited transparency into model decision-making. We present XKD-Dial, a progressive four-stage training pipeline for explainable, knowledge-grounded dialogue generation in a bilingual (English-Hindi) setting, comprising: (1) multilingual adaptation, (2) English dialogue SFT with citation grounding, (3) bilingual dialogue SFT, and (4) GRPO alignment with citation-aware rewards. We evaluate six models spanning encoder-decoder (250M-3B) and decoder-only (1B-7B) architectures at every pipeline stage. Our key contributions are: (i) three post-hoc explainability analyses - cross-attention alignment, Integrated Gradients attribution, and occlusion-based causal grounding - applied systematically across the training trajectory to reveal how citation behaviour is learned, not only whether it is learned; (ii) citation-grounded SFT reduces hallucination to 0.0% for encoder-decoder models from Stage 2 onward; (iii) the progressive pipeline prevents catastrophic forgetting while improving Hindi capabilities; (iv) smaller models match larger models on English after SFT; and (v) GRPO provides marginal improvement over well-designed SFT for structured citation tasks. We evaluate across six automatic metrics (BLEU, ROUGE, BERTScore, FactScore, Citation-F1, and hallucination rate).
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
