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

Vedant Pandya

Source-linked AI summary

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

Vedant Pandya

arXiv:2603.18911v1Mar 19, 2026cs.CLcs.AI

TL;DR

Knowledge-grounded dialogue lacks broad bilingual coverage, explicit claim-level citations, and transparent evidence of how models use retrieved passages. XKD-Dial addresses these gaps with progressive English-Hindi training and citation-aware alignment, achieving zero automatic NLI-based hallucination for encoder-decoder models from Stage 2 onward while retaining a marginal GRPO gain over SFT.

Problem
Knowledge-grounded dialogue remains limited by English-centric research, missing explicit citations for factual verification, and restricted transparency into grounding decisions.
Method
XKD-Dial trains bilingual dialogue models through multilingual adaptation, English citation-grounded SFT, bilingual SFT, and GRPO alignment with citation-aware rewards.
Results
0.0% hallucination is achieved under automatic NLI-based evaluation for encoder-decoder models from Stage 2 onward, while GRPO provides marginal improvement over SFT.
Takeaways & Limitations
Progressive training preserves English performance while improving Hindi capabilities, and well-designed SFT is a strong baseline for structured citation-grounded dialogue.
Takeaways & Limitations
Attention alignment analysis is limited to encoder-decoder models, and the Hindi examples are machine-translated rather than necessarily natural conversational data.

Abstract

from arXiv · show

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).

1 Introduction

Knowledge-grounded dialogue remains constrained by English-centric coverage, missing verifiable citations, and opaque grounding decisions. XKD-Dial addresses these gaps with progressive bilingual training, citation grounding, and explainability analysis.

Research gaps: Hindi knowledge-grounded dialogue lacks a standard benchmark, established training methodology, and systematic bilingual evaluation.The paper identifies limited Hindi dialogue data, morphological and word-order challenges, code-switching, and cross-lingual transfer as additional difficulties.
Research gaps: Explicit citation markers are needed to let users verify which knowledge passage supports each factual claim.The paper argues that retrieval alone does not identify claim-level evidence, limiting trust and transparency.
Research gaps: Systematic interpretability analysis is needed because correct citation markers may not reflect genuine conditioning on the cited passage.The paper proposes examining cross-attention, Integrated Gradients, and occlusion-based causal grounding across training.
Contributions: XKD-Dial uses four stages: multilingual adaptation, English citation-grounded SFT, bilingual dialogue SFT, and GRPO alignment.The pipeline incrementally adds bilingual, citation, dialogue, and reward-aligned capabilities.
Contributions: The study evaluates six models across encoder-decoder and decoder-only architectures from 250M to 7B parameters at every pipeline stage.This produces 30 evaluation runs for analyzing which stages contribute which capabilities.
Key findings: 0.0% hallucination is reached under automatic NLI-based evaluation for encoder-decoder models from Stage 2 onward.The paper presents citation-grounded SFT as a promising anti-hallucination strategy, while noting that human evaluation remains necessary.
Key findings: GRPO provides marginal improvement over SFT in the reported experimental configuration, while explicit citation training substantially reduces hallucination.The paper frames well-designed SFT as a strong baseline for structured citation-grounded tasks.
Explainability: Attention visualization and token attribution are applied to analyze how models attend to knowledge passages during generation.These analyses provide interpretability insights into citation-grounded dialogue behavior.

2 Related Work

Related work establishes knowledge-grounded dialogue, multilingual language modeling, reinforcement-learning alignment, and interpretability, but leaves Hindi citation-grounded dialogue insufficiently addressed. XKD-Dial combines these dimensions and analyzes grounding behavior during generation.

Knowledge-grounded dialogue: Knowledge-grounded dialogue research conditions responses on retrieved passages to improve informativeness and factual accuracy.Wizard of Wikipedia established the paradigm, while FaithDial targeted hallucination and faithfulness in grounded dialogue.
Knowledge-grounded dialogue: RAG, REALM, Fusion-in-Decoder, BlenderBot 2.0, and Atlas represent complementary retrieval, memory, aggregation, and scaling approaches.These systems advance retrieval-augmented generation but do not by themselves provide claim-level citation mechanisms.
Citation grounding: Existing grounded systems generally do not indicate which passage supports each claim, motivating inline citation generation for source verification.The paper positions citation mechanisms as a missing dimension in prior retrieval-conditioned systems.
Multilingual models: Multilingual pretrained models such as mT5, BLOOM, IndicBART, MuRIL, IndicTrans2, Flan-T5, LLaMA-3, and Gemma-2 expand language coverage and generation capability.The cited models provide multilingual or Indian-language foundations relevant to bilingual dialogue modeling.
Multilingual models: Hindi knowledge-grounded dialogue remains unexplored, with no existing work combining Hindi dialogue generation and citation grounding.XKD-Dial addresses this gap through a bilingual English-Hindi pipeline using progressive cross-lingual transfer.
Alignment: GRPO avoids a critic model by generating multiple outputs, ranking them by reward, and using relative rankings as the training signal.The paper applies GRPO with citation-aware rewards combining factual consistency, entity overlap, citation attribution, and hallucination penalties.
Explainability: Attention, Integrated Gradients, SHAP, and rationale extraction provide complementary approaches to neural interpretability.XKD-Dial applies attention visualization and token attribution to multilingual citation-grounded generation.
Positioning: XKD-Dial is presented as the first system to combine citation-grounded knowledge modeling, English-Hindi support, RL alignment, and explainability.The positioning is summarized in a comparison across key dimensions.

3 Methodology

XKD-Dial uses a progressive four-stage pipeline to build multilingual, citation-grounded dialogue capabilities incrementally. The task generates a cited response from an English or Hindi query and retrieved knowledge passages.

Pipeline design: The four-stage pipeline incrementally builds multilingual, citation-grounded dialogue capabilities.Each stage adds a specific capability while preserving skills learned earlier.
Task formulation: Given an English or Hindi query and retrieved passages K, the model generates a response r.The methodology frames response generation as conditioning on an external set of knowledge passages.

1. Is factually consistent with K,

The paper defines a four-stage pipeline for bilingual, citation-grounded dialogue, combining multilingual adaptation, supervised fine-tuning, and citation-aware GRPO alignment. It also evaluates post-hoc explainability methods to assess whether generated citations correspond to supporting knowledge.

Stage 1: Multilingual Adaptation: Stage 1 uses bidirectional English–Hindi translation training to build bilingual representations while limiting adaptation to one epoch to reduce catastrophic interference.Encoder-decoder models use parallel sentence pairs, while decoder-only models use instruction-formatted translation with causal language modeling.
Stage 2: English Citation-Grounded SFT: Stage 2 applies English supervised fine-tuning on structured prompts and responses containing inline citations to learn dialogue generation, citation attachment, and knowledge grounding.The citation format is intended to make claims supportable by numbered knowledge references.
Stage 3: Bilingual Dialogue SFT: Stage 3 mixes English and Hindi cited dialogue examples, weighting Hindi at 0.6 and English at 0.4 to improve Hindi while replaying English.The paper reports that citation formatting learned in English transfers effectively to Hindi without relearning the citation pattern.
Stage 4: GRPO Alignment: Stage 4 uses GRPO with a citation-aware composite reward that combines multiple quality signals and strongly penalizes fabricated citation indices.The hallucination penalty has weight 10.0, the highest reward weight described.
Explainability Module: Three post-hoc analyses inspect citation behaviour through cross-attention, Integrated Gradients, and minimal sufficient knowledge-token extraction.The explainability module does not affect training and is used to test whether citations align with genuine internal grounding.

4 Experimental Setup

The experiments use a balanced bilingual corpus assembled from three knowledge-grounded dialogue benchmarks, with English data translated to Hindi while preserving citations. Six models are evaluated across six metrics at five training stages.

Datasets: The bilingual corpus combines DSTC9, FaithDial, and Wizard of Wikipedia, with English portions translated to Hindi using IndicTrans2.The source benchmarks cover task-oriented, faithful, and open-domain knowledge-grounded dialogue.
Datasets: Citation markers are preserved during English-to-Hindi translation through regex-based preprocessing and postprocessing.This keeps citation references available in the translated dialogue data.
Training Configuration: Training uses AdamW, BFloat16 mixed precision, cosine learning-rate scheduling with warmup, and an effective batch size of 64 via gradient accumulation.
Evaluation: Evaluation covers BLEU, ROUGE-1/ROUGE-L, BERTScore, FactScore, Hallucination Rate, Citation F1, and Has Citation.
Evaluation: Each of six models is evaluated at baseline and Stages 1–4, producing 30 evaluation runs with greedy decoding and up to 128 generated tokens.Metrics are computed separately for English and Hindi.

5 Results and Analysis

Across training stages, citation-grounded SFT sharply improves evaluation results, eliminates hallucinations for encoder-decoder models, and transfers citation formatting to Hindi. The results also reveal model-specific failures, including XL generation collapse and LLaMA-1B’s separation of hallucination control from citation learning.

Overall results: Stage 2 produces a sharp phase transition: overall metrics improve dramatically, with hallucination reaching exactly 0.0 for encoder-decoder models.The progression is reported across BLEU, ROUGE-L, BERTScore, FactScore, Citation F1, and hallucination rate.
English results: After Stage 2, Base and Large reach identical English BLEU (0.172), Citation F1 (0.980), and BERTScore (0.889).This indicates that the 250M Base model matches the larger model on these structured English metrics after SFT.
English results: Stage 1 preserves English performance within ±0.002 for encoder-decoder models, but LLaMA-1B’s hallucination rate rises from 13.5% to 66.5%.The increase is described as a 4.9× spike and contrasts with encoder-decoder stability.
Hindi results: Stage 3 raises Base Hindi ROUGE-1 from 0.481 to 0.691 (+0.210), the largest single-stage improvement reported.The result supports the importance of bilingual SFT with Hindi examples.
Hindi results: English-only Stage 2 increases Base Hindi Citation F1 from 0.485 to 0.718, showing that citation formatting transfers across languages.The authors characterize citation formatting as a language-agnostic structural skill.
Failure modes: XL produces empty outputs after Stage 2 but recovers in Stage 3, reaching overall BLEU 0.096, Citation F1 0.898, and BERTScore 0.765.The collapse is attributed to an aggressive 2 × 10^-5 learning rate, while continued bilingual training rescues the checkpoint.
Failure modes: Teacher-forced loss can appear healthy while autoregressive generation collapses, so generation-time validation is necessary alongside loss monitoring.The XL Stage 2 checkpoint had loss 1.642 despite producing empty outputs.
Failure modes: LLaMA-1B reaches 0.0% English hallucination from Stage 2 while English Citation-F1 remains 0.000; by Stage 3, Hindi Citation-F1 reaches 0.783.The model avoids hallucination without learning English citation markers, demonstrating separable objectives and language-selective citation learning.

5.6 Stage 4 (GRPO) Effectiveness

Stage 4 GRPO adds little measurable benefit over Stage 3 SFT in this structured citation task, while reward trajectories reveal instability for most models. Explainability analyses show that citation format learning and genuine source grounding diverge across architectures.

GRPO effectiveness: GRPO deltas are negligible for encoder-decoder models, while Mistral shows only marginal gains in Citation F1 and FactScore.The comparison covers changes from Stage 3 to Stage 4.
GRPO dynamics: Reward declines from best to final for most models, suggesting unstable GRPO training; Gemma-2-2B is the exception with identical best and final rewards.Gemma reaches an absolute reward of 3.40 with zero decline, while Mistral reaches 2.889 with a decline of −0.457.
Training outcomes: Mistral-7B achieves the lowest loss across stages, with Stage 3 loss 0.303 versus XL’s 0.895.The Stage 2 → Stage 3 loss drop is consistent across models, indicating improved fit to the dialogue distribution after bilingual SFT.
Explainability analyses: Cross-attention alignment peaks at Stage 2 for Flan-T5 models, while XL’s Stage 2 alignment of 0.000 corroborates its generation collapse.After recovery at Stage 3, XL reaches 0.026 alignment, above Base at 0.019 and Large at 0.022.
Explainability analyses: Training increases saliency entropy and decreases concentration for Flan-T5 models, indicating broader attribution across knowledge passages.XL’s Stage 2 NaN saliency independently confirms that its generation collapse prevented meaningful gradient signals from the output.
Explainability analyses: Occlusion analysis shows encoder-decoder grounding improvements, but both decoder-only models reach 0.000 causal grounding despite retaining citation markers.Mistral falls from 0.767 baseline grounding to 0.000 with Citation-F1 0.772, while Gemma falls from 0.733 to 0.000 with Citation-F1 0.903 at Stage 3.

6 Discussion

The discussion finds architecture- and model-dependent effects across hallucination reduction, language quality, citation learning, and recovery from collapse. It also cautions that GRPO conclusions are limited by the single experimental configuration and that automatic comparisons require careful interpretation.

6.1 Citation-Grounded SFT as Anti-Hallucination: Encoder-decoder hallucination reaches 0.0% after Stage 2, while Mistral-7B stabilizes at 0.014 after Stage 3.LLaMA-3.2-1B also reaches 0.0% English hallucination, but generates no citation markers.
6.1 Citation-Grounded SFT as Anti-Hallucination: LLaMA-3.2-1B suppresses hallucination through conservative generation rather than citation grounding, showing that zero hallucination and citation grounding are separable outcomes.Its English Citation-F1 remains 0.000 from Stage 2 onward.
6.2 SFT Sufficiency for Structured Tasks: GRPO provides marginal improvement over SFT in this configuration, possibly because SFT already satisfies the structured task’s clear quality criteria.The authors explicitly avoid generalizing beyond this point in the hyperparameter space.
6.3 Evaluation Metrics: Hindi BLEU remains near-zero for encoder-decoder models despite ROUGE-1 up to 0.691 and BERTScore up to 0.624, exposing a metric limitation.Decoder-only BLEU is higher but still underestimates quality relative to ROUGE-1 and BERTScore.
6.4 Model Scale: Base and Large match on English after Stage 2 at BLEU 0.172, BERTScore 0.889, and Citation F1 0.980.The reported result suggests no English benefit from additional capacity for this constrained task.
6.4 Model Scale: Progressive training narrows the baseline Hindi BERTScore gap between Large and Base from 0.617 versus 0.221 to 0.615 versus 0.624 after Stage 3.The discussion attributes the initial gap to greater Hindi exposure in the larger pretraining corpus.
6.5 Generation Collapse: Flan-T5-XL recovers from Stage 2 collapse at Stage 3, while its collapse is attributed to an aggressive 2 × 10−5 learning rate.The recovered model achieves performance comparable to Base and Large.

7 Conclusion and Future Work

XKD-Dial establishes a progressive bilingual citation-grounded dialogue pipeline, with evaluation showing reduced hallucination, retained English performance, improved Hindi capability, and strong SFT baselines. The authors identify human evaluation, broader language coverage, and improved GRPO as future directions.

Key Findings: 0.0% hallucination was achieved by encoder-decoder models after Stage 2 under automatic NLI-based evaluation, while Mistral-7B reached 0.010–0.014.LLaMA-3.2-1B also reached 0.0% English hallucination without producing citation markers, separating hallucination elimination from citation learning.
Key Findings: English metrics remained stable while Hindi capabilities improved substantially through Stages 1–3.The result supports the pipeline’s incremental skill-composition approach.
Key Findings: GRPO provided marginal improvement over SFT in the experimental configuration, leaving well-designed SFT as a competitive baseline for citation-grounded dialogue.The authors note that reinforcement learning under other configurations requires further investigation.
Key Findings: A 250M-parameter model achieved identical English performance to a 780M model after Stage 2, with implications for cost-effective deployment.This comparison concerns English performance after supervised fine-tuning.
Key Findings: Flan-T5-XL experienced generation collapse at Stage 2 but recovered fully at Stage 3.The finding highlights the need for generation-time validation during continued training.
Future Work: Future work targets improved GRPO, human evaluation, additional Indian languages, explainability-guided training, and dataset- or language-specific training studies.The proposed extensions include reducing the KL penalty, increasing training steps, graded rewards, and separate corpus or language tracks.

Appendix A: LLaMA-3.2-1B Generated Output Examples

The appendix presents representative Stage 3 LLaMA-3.2-1B outputs against citation-grounded references, revealing different citation behavior across English and Hindi.

Output Comparison: English outputs were factually correct but contained no citation markers despite access to the numbered knowledge passage.This pattern is reported for representative Stage 3 outputs.
Output Comparison: Hindi outputs correctly generated integrated [N] citation markers in natural-language phrasing.The table compares these outputs with expected citation-grounded references.

Appendix B: Figures

The appendix figures document XKD-Dial’s architecture, data and optimization schedules, and training-loss behavior across models and stages.

System Architecture: User queries undergo language identification and optional translation before retrieval, after which a context builder combines the query with top-k knowledge snippets.A multilingual generator produces the citation-grounded response, followed by explainability analysis.
Data Distribution: The data distribution covers DSTC9, FaithDial, and Wizard of Wikipedia across English and Hindi train, validation, and test partitions.The figure summarizes source-corpus and language splits.
Optimization: Cosine decay with warmup is used for SFT, while GRPO uses a lower learning rate with linear warmup.The schedules differ across training objectives.
Validation Loss: Mistral-7B achieves the lowest validation loss throughout, while Gemma-2-2B reaches 0.651 at Stage 3.Flan-T5-Large uses training loss as a Stage 1 fallback because evaluation loss was not logged.
GRPO Reward: Gemma-2-2B reaches the highest absolute GRPO reward, 3.40, without decline, whereas most other models regress from best to final reward.The comparison covers all six models during Stage 4.
Training Loss: Training loss converges consistently across architectures during Stage 2 English SFT and Stage 3 bilingual SFT.The figure tracks loss progression across all training stages.
Language-Specific Loss: Hindi loss starts higher during Stage 3 but decreases rapidly until converging with English loss.The figure isolates language-specific training loss.
Policy Stability: A KL penalty of β = 0.04 constrains policy drift between the policy and reference models during GRPO.The figure reports the policy-reference KL divergence over training.

Appendix C: Additional Evaluation Figures

Additional figures characterize GRPO reward distributions, best-versus-final reward changes, and policy divergence for Mistral-7B.

Reward Distributions: Mistral-7B shows a compact, high-reward distribution centered around 3.0, while encoder-decoder models have wider spreads and lower medians.The violin plot compares reward distributions by model.
Reward Stability: Flan-T5-XL exhibits the largest best-to-final reward drop at −1.30, while Mistral-7B has the smallest drop at −0.46.All models show reward decline from best to final.
Policy Divergence: Mistral-7B’s negative KL divergence stabilizes around −2.0 after 300 steps despite the KL penalty β = 0.04.The figure tracks divergence from the reference policy during GRPO.