Progressive Training for Explainable Citation-Grounded Dialogue: Reducing Hallucination to Zero in English-Hindi LLMs

1. Problem Statement

The paper addresses three critical limitations in current Knowledge-Grounded Dialogue (KGD) systems:

1. Monolingual Bias: Existing research focuses almost exclusively on English, leaving low-resource languages like Hindi (spoken by 600M+ people) without standardized benchmarks or training methodologies for bilingual, knowledge-grounded dialogue.
2. Lack of Verifiable Citations: While Retrieval-Augmented Generation (RAG) systems retrieve knowledge, they rarely generate explicit citation markers (e.g., "[1]"). This prevents users from verifying factual claims against source passages, undermining trust.
3. Model Opacity: Even when models generate correct responses, the decision-making process is opaque. It is unclear if a model genuinely conditions its output on a specific knowledge passage or merely mimics citation formatting without semantic grounding.

2. Methodology: XKD-Dial Pipeline

The authors propose XKD-Dial, a progressive four-stage training pipeline designed to incrementally build bilingual, citation-grounded capabilities while preventing catastrophic forgetting.

• Stage 1: Multilingual Adaptation

  • Goal: Adapt pretrained models to bilingual English-Hindi representations.
  • Method: Translation training on parallel corpora (IIT Bombay).
  • Strategy: Limited to one epoch to establish bilingual alignment without overwriting pre-trained instruction-following capabilities.

• Stage 2: English Dialogue SFT (Supervised Fine-Tuning)

  • Goal: Teach the model to generate responses with explicit citation markers linked to knowledge passages.
  • Method: SFT on English knowledge-grounded dialogue data where outputs follow a strict format (e.g., "According to [1]...").
  • Mechanism: The citation format acts as an implicit anti-hallucination constraint, forcing the model to associate claims with numbered references.

• Stage 3: Bilingual Dialogue SFT

  • Goal: Extend capabilities to Hindi while preserving English performance.
  • Method: SFT on a weighted mixture of English and Hindi dialogue data (40% EN, 60% HI).
  • Key Insight: Citation formatting learned in Stage 2 transfers cross-lingually; the model applies the [1], [2] pattern to Hindi without relearning the mechanics.

• Stage 4: GRPO Alignment

  • Goal: Further align the model with citation quality and factual consistency.
  • Method: Group Relative Policy Optimization (GRPO) using a composite reward function.
  • Reward Components: Factual consistency (NLI-based), entity overlap, citation attribution accuracy, fluency, and heavy penalties for hallucination and incorrect citations.

3. Experimental Setup

• Models: Six models spanning two architectures and scales (250M to 7B parameters):
  • Encoder-Decoder: Flan-T5 (Base, Large, XL).
  • Decoder-Only: LLaMA-3.2-1B, Gemma-2-2B, Mistral-7B.
• Datasets: A bilingual dataset constructed by translating three English benchmarks (DSTC9, FaithDial, Wizard of Wikipedia) into Hindi using IndicTrans2.
• Explainability Analysis: Three post-hoc methods applied across the training trajectory:
  1. Cross-attention visualization (for encoder-decoder).
  2. Integrated Gradients (token attribution).
  3. Occlusion-based causal grounding (testing if citations disappear when source text is removed).

4. Key Results

A. Hallucination Reduction

• Encoder-Decoder Models: Citation-grounded SFT (Stage 2) reduced the hallucination rate to 0.0% for Flan-T5 Base and Large. This zero-hallucination state was maintained through Stages 3 and 4.
• Decoder-Only Models: Mistral-7B saw a significant drop in hallucination (from 7.8% to ~1.4%).
• The LLaMA-3.2-1B Anomaly: This model achieved 0.0% English hallucination after Stage 2 but failed to generate any citation markers (Citation-F1 = 0.000). It achieved "zero hallucination" by adopting a conservative, non-committal generation strategy rather than grounding claims in sources.

B. Architecture and Scaling Effects

• Small vs. Large Models: After Stage 2 SFT, the 250M Flan-T5-Base achieved identical English performance (BLEU, Citation-F1, BERTScore) to the 780M Flan-T5-Large, suggesting that for structured citation tasks, small models are sufficient.
• Generation Collapse: Flan-T5-XL (3B) suffered a "generation collapse" at Stage 2 (producing empty strings) due to an aggressive learning rate. However, it fully recovered at Stage 3 (Bilingual SFT), demonstrating the resilience of continued training.
• Decoder-Only Limitations: While decoder-only models learned citation formats, they failed causal grounding. Occlusion tests showed that for Mistral-7B and Gemma-2-2B, removing the source text did not change the citation markers (0.000 causal grounding), implying they learned the syntax of citations without the semantic dependency on the source.

C. GRPO Effectiveness

• Marginal Gains: GRPO (Stage 4) provided negligible improvements over well-designed SFT. For encoder-decoder models, metrics remained static. For decoder-only models, gains were marginal (e.g., Mistral-7B Citation-F1 +0.004).
• Implication: For highly structured tasks like citation-grounded dialogue, SFT may be sufficient, and RL alignment adds little value unless the reward signal is significantly more granular.

D. Explainability Findings

• Encoder-Decoder: Cross-attention weights increased significantly at Stage 2, showing the model actively attending to cited knowledge tokens when generating markers.
• Decoder-Only: Lacked cross-attention mechanisms. Gradient saliency analysis revealed that while they produced citations, the attention was not causally grounded in the source text.

5. Significance and Contributions

1. First Bilingual Citation Pipeline: XKD-Dial is the first systematic framework for English-Hindi knowledge-grounded dialogue with explicit citations.
2. Zero Hallucination via SFT: Demonstrates that explicit citation formatting in SFT can reduce hallucination to 0.0% for encoder-decoder models, offering a viable alternative to complex RL pipelines.
3. Decoupling Hallucination and Citations: The study reveals that a model can eliminate hallucinations without learning to cite (by being vague), highlighting that "zero hallucination" metrics alone are insufficient without checking for citation grounding.
4. Architecture-Specific Grounding: Identifies a critical gap where decoder-only models learn citation syntax but fail semantic grounding (causal dependency on source), whereas encoder-decoder models maintain true grounding via cross-attention.
5. Cost-Effective Deployment: Shows that small models (250M) can match large models (780M+) in structured tasks after proper SFT, enabling cheaper deployment.

6. Conclusion

The paper establishes that a progressive training pipeline effectively builds bilingual, explainable dialogue systems. While SFT is highly effective at reducing hallucinations and teaching citation formats, the study warns that citation accuracy metrics (like Citation-F1) can be misleading if not paired with causal grounding tests (like occlusion), especially for decoder-only architectures. Future work should focus on human evaluation and refining RL strategies to ensure semantic grounding rather than just syntactic compliance.