Progressive Citation Grounded Dialogue

Progressive Citation-Grounded Dialogue

Ranked Findings

Finding 1 (Highest Impact): Citation-Grounded SFT Eliminates Hallucination

• Stages 2+ achieve 0.0% hallucination (encoder-decoder models under NLI-based evaluation)
• Structured format (citations to source passages) forces models to ground outputs
• No downstream GRPO alignment needed for hallucination elimination

Finding 2: Encoder-Decoder Architecture Superior to Decoder-Only

• Flan-T5 models: consistent 0.0% hallucination
• LLaMA-3.2-1B: language-selective citation failure (English works, Hindi fails)
• Mistral: larger decoders (7B) partially recover but never reach encoder-decoder parity
• Implication: model architecture matters as much as training procedure

Finding 3: Smaller Models Can Match Larger Ones with Structured Training

• Flan-T5-250M reaches English metrics of Flan-T5-780M after Stage 2
• Structured citation task requires <250M capacity if properly formatted
• Training recipe supersedes raw model size for hallucination control

Finding 4: GRPO Alignment Provides Marginal Gains

• Well-tuned SFT already achieves strong performance
• GRPO adds <2% improvement on citation-F1 and factuality scores
• Resource cost (reward modeling, RL training) not justified for this task
• Implication: structured generation tasks may not need RL fine-tuning

Finding 5: Cross-Lingual Citation Transfer is Automatic

• Models trained on English citation format automatically apply format to Hindi
• No need for language-specific instruction tuning for citation structure
• Citation formatting is learned as abstract pattern, not language-dependent

Decision Checklist for Practitioners

1. Choose Encoder-Decoder Foundation: Start with encoder-decoder models (Flan-T5, mBART)

   • ✓ Encoder-decoder models exhibit 0.0% hallucination
   • ✗ Avoid decoder-only for critical applications unless scaling to 7B+

2. Prepare Citation-Grounded Data: Format SFT data with explicit source attributions

   • Format: "[Response text] [CITE: source_passage_id, source_passage_id]"
   • Verify source passages are included in context; no floating citations
   • Target: 1,000+ citation-grounded training examples minimum

3. Run Stage 2 (English Dialogue SFT) First:

   • Monitor for generation collapse (if occurs, will appear at Stage 2)
   • Verify 0.0% hallucination before proceeding to multilingual training
   • If collapse happens: reduce learning rate, increase training steps

4. Add Stage 3 (Bilingual SFT) with Weighted Sampling:

   • 60% target-language (Hindi), 40% English to prevent English forgetting
   • Verify citation format transfers to target language automatically
   • Accept minor performance drop in target language; citation format is preserved

5. Skip Stage 4 (GRPO) Unless Additional Metrics Required:

   • Hallucination already eliminated by Stage 2
   • GRPO only helpful if optimizing citation coverage (citing every relevant passage)
   • For factuality-only tasks, SFT sufficient

6. Evaluation Strategy:

   • Primary: NLI-based hallucination detection (0.0% target)
   • Secondary: Citation-F1 (precision and recall of citations)
   • Tertiary: BLEU, ROUGE (downstream task quality)
   • Sample evaluation: 100+ conversations with human verification of facts

Conditions for Success

Necessary Conditions:

1. Source Knowledge Must Be Available: If knowledge sources aren't in model context, citation instructions won't work; system needs explicit passages to cite
2. Structured Format Compliance: Models must follow citation marker format (e.g., "[CITE: 1, 2]"); test format consistency during training
3. Sufficient Training Data: Minimum 1,000 citation-grounded dialogues; below this, models revert to ungrounded generation
4. Encoder-Decoder Architecture: Decoder-only models struggle with structured citation; if using decoder-only, scale to 7B+ and accept partial failure modes

Sufficient Conditions (Optional Enhancements):

1. Multilingual Pretraining: mBART or multilingual BERT helps Stage 1; monolingual models work but need more adaptation data
2. GRPO Alignment: Not necessary for hallucination elimination, but improves citation precision (cite only most relevant passages)
3. Explicit Causal Grounding: Use Integrated Gradients or occlusion analysis to verify citations actually influence model outputs (not just format compliance)

Notable Phenomena

Phenomenon 1: Generation Collapse and Recovery

• Flan-T5-XL experienced complete generation failure at Stage 2 (outputting only "[CITE: X]")
• Recovered fully during Stage 3 bilingual training (bilingual context restabilized representations)
• Implication: citation format can destabilize some models; multilingual exposure helps recovery

Phenomenon 2: Decoder-Only Dissociation

• Larger decoder-only models (LLaMA-3.2-7B, Mistral-7B) learn citation FORMAT without causal grounding
• Models output correct citations but base responses on other factors, not cited passages
• Contrast: Encoder-decoder models inherently ground via cross-attention; citations reflect actual decision-making
• Implication: architecture enables or disables genuine grounding, not just format learning

Phenomenon 3: Cross-Lingual Transfer Without Explicit Instruction

• Models never see Hindi citation examples (Stage 2 English-only); automatically apply format to Hindi
• Transfer happens in Stage 3 without additional citation-specific instructions
• Suggests citation formatting learned as abstract structural pattern
• Opportunity: extend to zero-shot languages (untested in this work)

Training Pipeline Details

Stage 1 (Multilingual Adaptation):

• English-Hindi parallel corpus (existing public datasets)
• Masked language modeling or back-translation
• Duration: 1-2 epochs

Stage 2 (English Dialogue SFT):

• Citation-grounded dialogues with English source passages
• Supervised fine-tuning with cross-entropy loss
• Monitor for generation collapse; reduce learning rate if detected
• Duration: 5-10 epochs or until validation plateau

Stage 3 (Bilingual SFT):

• Extend to Hindi with weighted sampling (60% Hindi, 40% English)
• Same citation format as Stage 2
• Higher learning rate than Stage 2 to prevent catastrophic forgetting
• Duration: 3-5 epochs

Stage 4 (GRPO, Optional):

• Group Relative Policy Optimization with citation-aware rewards
• Reward: bonus for citations covering diverse source passages, penalty for over-citing
• Only if optimizing for citation coverage/diversity
• Duration: 1-3 epochs

Practical Implications

• Hallucination Elimination at Scale: Citation-grounded training transfers to longer conversations and complex topics
• Multilingual Scalability: Approach tested on English-Hindi; architecture generalizes to other language pairs
• Production Deployment: 0.0% hallucination under evaluation metrics suggests production-ready factuality
• User Trust: Explicit citations enable users to verify facts and trace reasoning