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).
Figures & tables
System
Cite
Hindi
RL
XAI
Wizard of Wikipedia
×
×
×
×
FaithDial
×
×
×
×
BlenderBot 2.0
×
×
×
×
RAG
×
×
×
×
Atlas
×
×
×
×
InstructGPT
×
×
✓
×
Table 1: Comparison with related work across key dimensions. XKD-Dial is the first to address all four dimensions simultaneously.
Model
Type
Params
Layers
Flan-T5-Base
Enc-Dec
250M
12+12
Flan-T5-Large
Enc-Dec
780M
24+24
Flan-T5-XL
Enc-Dec
3B
24+24
LLaMA-3.2-1B
Dec-Only
1B
16
Gemma-2-2B
Dec-Only
2B
26
Mistral-7B
Dec-Only
7B
32
Table 2: Model architectures used in our study.
Sym.
Component
w
rfact
Factual consistency (NLI)
5.0
rent
Entity overlap
3.0
rattr
Citation attribution
1.5
rflu
Fluency proxy
1.0
rlen
Length penalty
− 0.1
rhal
Hallucination penalty
− 10.0
Table 3: GRPO reward function components with weights.
Split
Total
EN
HI
EN%
Train
135,000
71,957
63,043
53.3
Val
7,500
4,014
3,486
53.5
Test
7,500
4,029
3,471
53.7
Table 4: Dataset statistics. All splits maintain a balanced English–Hindi distribution.
Param.
Base
Large
XL
Mistr.
Stages 1–3: Supervised Fine-Tuning
LR
5e-5
3e-5
3e-5
1e-5
Batch
8
4
2
2
Grad.Acc.
4
8
16
16
Eff.Batch
64
64
64
64
Epochs
3
3
3
3
Table 5: Training hyperparameters. Effective batch size is held constant at 64 via gradient accumulation.
Model
Stage
BLEU
ROUGE-1
ROUGE-L
FactScore
Cit-F1
Halluc.
BERTScore
Encoder-Decoder Models
Flan-T5-Base (250M)
Baseline
0.004
0.201
0.201
0.059
0.673
0.005
0.551
Stage 1
0.005 ∼
0.238 ↑
0.237 ↑
0.056 ∼
0.738 ↑
0.028 ↑
0.611 ↑
Stage 2
0.094 ↑
0.412 ↑
0.388 ↑
0.106 ↑
0.859 ↑
0.000 ↓
0.739 ↑
Stage 3
0.092 ∼
0.507 ↑
0.483 ↑
0.096 ∼
0.902 ↑
0.000 ∼
0.766 ↑
Stage 4
0.092 ∼
0.507 ∼
0.483 ∼
0.098 ∼
0.902 ∼
0.000 ∼
0.766 ∼
Table 6: Overall evaluation results across all stages. Best result per model is bolded . “ † ” indicates generation collapse (empty outputs). Hallucination rate of 0.0 from Stage 2 onward is highlighted.
Figure 7
Model
Stage
BLEU
ROUGE-1
ROUGE-L
FactScore
Cit-F1
Halluc.
BERTScore
Encoder-Decoder Models
Base (250M)
Baseline
0.007
0.100
0.100
0.039
0.960
0.005
0.836
Stage 1
0.007 ∼
0.101 ∼
0.101 ∼
0.040 ∼
0.956 ∼
0.004 ∼
0.835 ∼
Stage 2
0.172 ↑
0.354 ↑
0.308 ↑
0.121 ↑
0.980 ↑
0.000 ↓
0.889 ↑
Stage 3
0.170 ∼
0.349 ∼
0.304 ∼
0.120 ∼
0.980 ∼
0.000 ∼
0.889 ∼
Stage 4
0.170 ∼
0.349 ∼
0.304 ∼
0.124 ∼
0.980 ∼
0.000 ∼
0.889 ∼
Table 7: English evaluation results. After Stage 2 SFT, Base (250M) and Large (780M) converge to identical performance across all metrics. Citation F1 reaches 0.980 for both models.
Model
Stage
BLEU
ROUGE-1
ROUGE-L
FactScore
Cit-F1
Halluc.
BERTScore
Encoder-Decoder Models
Base (250M)
Baseline
0.000
0.318
0.318
0.082
0.340
0.005
0.221
Stage 1
0.003 ∼
0.396 ↑
0.395 ↑
0.075 ∼
0.485 ↑
0.056 ↑
0.351 ↑
Stage 2
0.003 ∼
0.481 ↑
0.481 ↑
0.088 ↑
0.718 ↑
0.000 ↓
0.565 ↑
Stage 3
0.001 ∼
0.691 ↑
0.691 ↑
0.069 ↓
0.812 ↑
0.000 ∼
0.624 ↑
Stage 4
0.001 ∼
0.690 ∼
0.690 ∼
0.069 ∼
0.811 ∼
0.000 ∼
0.623 ∼
Table 8: Hindi evaluation results. Stage 1 builds the foundation, Stage 2 transfers citation skills, and Stage 3 provides the largest Hindi improvement. Note: Hindi BLEU is near-zero due to morphological variation (see Section 6 ).
Figure 3: Per-language BLEU progression across training stages. English BLEU (left) shows convergence of all models after Stage 2 SFT, while Hindi BLEU (right) remains near-zero for encoder-decoder models due to morphological variation (Section 6.3 ). Gemma-2-2B achieves the highest Hindi BLEU (0.155) after Stage 3, followed by Mistral-7B (0.070).
Metric
Stage 3
Δ
Encoder-Decoder
Base
Cit-F1
0.902
0.000
Halluc.
0.000
0.000
BERTScore
0.766
0.000
FactScore
0.096
+ 0.002
Large
Cit-F1
0.896
− 0.001
Table 9: Change from Stage 3 to Stage 4 (GRPO). Deltas are negligible for encoder-decoder models. Mistral shows marginal gains in Citation F1 and FactScore.
Model
Best
Final
Δ
Enc-Dec
Base (250M)
1.040
0.648
− 0.391 ↓
Large (780M)
0.868
0.157
− 0.711 ↓
XL (3B)
1.782
0.482
− 1.300 ↓
Dec-Only
LLaMA (1B)
3.08
2.51
− 0.57 ↓
Gemma (2B)
3.40
3.40
0.00
Mistral (7B)
2.889
2.433
− 0.457 ↓
Table 10: GRPO reward trajectory during Stage 4 training. Encoder-decoder and decoder-only models (LLaMA, Mistral) show reward decline from best to final, suggesting training instability. Gemma-2-2B is the exception: its best reward equals its final reward, indicating reward improvement continued to the last step.
Figure 13
Model
Stage
Steps
Best Loss
Encoder-Decoder
Base
Stage 1
3,906
1.552
Stage 2
7,500
1.916
Stage 3
4,000
0.945
Large
Stage 1
3,906
1.439
Stage 2
8,973
1.776
Stage 3
4,000
0.912
Table 11: Training summary showing total steps and best validation loss per stage. Mistral-7B consistently achieves the lowest loss across all stages.
Model
Base
S1
S2
S3
S4
Flan-T5-Base
0.017
0.022
0.035
0.019
0.019
Flan-T5-Large
0.030
0.032
0.037
0.022
0.022
Flan-T5-XL
0.003
0.004
0.000 †
0.026
0.026
Gemma-2-2B
N/A (decoder-only)
LLaMA-3.2-1B
N/A (decoder-only)
Mistral-7B
N/A (decoder-only)
Table 12: Mean cross-attention alignment with cited knowledge tokens (encoder-decoder models only). Decoder-only models show 0.000 by architecture. Stage 2 collapse for XL is marked with † .
Figure 6: Representative cross-attention heatmap for Flan-T5-Base (Stage 3). Output tokens (y-axis) attend to input knowledge tokens (x-axis). Brighter cells indicate higher attention weight. Citation-relevant tokens receive disproportionately high attention.
Baseline
Stage 2
Stage 4
Model
Entr.
Conc.
Entr.
Conc.
Entr.
Conc.
Base
4.36
0.378
4.44
0.343
4.50
0.326
Large
4.45
0.327
4.48
0.322
4.49
0.320
XL
4.38
0.360
NaN
NaN
4.53
0.306
Table 13: Gradient saliency entropy and concentration for Flan-T5 encoder-decoder models across training stages. Higher entropy = more diffuse attribution; lower concentration = less dominated by top tokens. XL Stage 2 NaN confirms generation collapse.
Model
Baseline
Stage 1
Stage 3
Stage 4
Flan-T5-Base
0.647
0.691
0.889
0.889
Flan-T5-Large
0.656
0.630
0.909
0.909
Flan-T5-XL
0.960
0.964
0.808
0.808
LLaMA-3.2-1B
0.696
0.544
0.722
0.632
Gemma-2-2B ‡
0.733
0.000
0.000
0.000
Mistral-7B †
0.767
0.797
0.000
0.000
Table 14: Occlusion causal grounding: fraction of citations that disappear when source is removed. Higher = stronger causal grounding. Both Mistral-7B and Gemma-2-2B drop to 0.000 post-training, indicating decoder-only models learn citation format without source-conditioned grounding.
L
Query (excerpt)
Expected Output
LLaMA-3.2-1B Output
Cit.
EN
Can you bring your own liquor to Pizza Hut Cherry Hinton?
As noted in [1], No, outside beverages are not allowed.
No, you cannot bring your own liquor at Pizza Hut Cherry Hinton. Do you have any other questions? Would you like to make a reservation? …
0
EN
How long did California Love stay on Billboard Hot 100?
Based on [1], yeah, stayed on the billboard hot 100’s for two weeks straight!
It was one of 2Pac’s most popular singles, reaching number one on the Billboard Hot 100 for two weeks. It was a double A-side single …
0
EN
Restaurant booking confirmation
According to [1], booking was successful. Ref: ENOALFAV .
Your reservation is all set. The reference number is 1ZQD0Z1A . …
0
HI
[ Query about wheelchair accessibility at Prezzo ]
[2] \devanagarifont में कहा गया है कि हां, प्रेज़ो में व्हीलचेयर उपलब्ध है।
[2] \devanagarifont के आधार पर, हां, प्रेज़ो में व्हीलचेयर उपलब्ध है इसलिए उन्हें आपकी सहायता करने में सक्षम होना चाहिए।
[2] ✓
HI
[ Query about the color black ]
[1] \devanagarifont में उल्लेख है कि काले रंग के बारे में, यह पूरी तरह से दिखाई देने वाले प्रकाश को अवशोषित करता है।
[1] \devanagarifont में कहा गया है कि मुझे नहीं पता था कि यह था, काले रंग के बारे में…
[1] ✓
Table 15: Representative LLaMA-3.2-1B Stage 3 outputs vs. expected references. EN outputs contain zero citation markers despite the model having access to the numbered knowledge passage. HI outputs correctly produce [N] markers.
Figure 20Figure 21Figure 22Figure 23Figure 24
Figure 17: KL divergence (policy vs. reference) for Mistral-7B during GRPO. The negative KL indicates the policy diverges from the reference, stabilizing around − 2.0 after 300 steps despite the KL penalty ( β=0.04 ).
Full-duplex spoken dialogue systems can model natural conversational behaviours such as interruptions, overlaps, and backchannels, yet such systems remain largely unexplored for Indian languages. We present the first open, reproducible full-duplex spoken dialogue system for Hindi by adapting Moshi, a state-of-the-art duplex speech architecture, using a custom Hindi tokeniser and training on 26,000 hours of real spontaneous conversations collected from 14,695 speakers with separate speaker channels, enabling direct learning of turn-taking and overlap patterns from natural interactions. To support Hindi text generation, we replace the original English tokeniser and reinitialise text-vocabulary-dependent parameters while retaining the pre-trained audio components. We propose a two-stage training recipe -- large-scale pre-training followed by fine-tuning on 1,000 hours of conversational data. Evaluation through the prompted dialogue continuation paradigm with both automatic metrics and human judgments demonstrates that the resulting model generates natural and meaningful full-duplex conversational behaviour in Hindi. This work serves as a first step toward real-time duplex spoken dialogue systems for Hindi and other Indian languages.
LLMs deployed multilingually are often audited via English explanations for non-English inputs. We evaluate extractive explanations ''where the model identifies input token spans as evidence alongside a generated rationale'' and uncover a systematic trade-off: English-pivot explanations can achieve higher span agreement with human rationales while their evidence becomes less causally grounded in the model's prediction, as measured by both comprehensiveness and sufficiency. Across 3 tasks, 5languages, and 2multilingual LLM families, we find that English explanations frequently produce fluent but loosely anchored rationales, with comprehensiveness degrading by up to 5.7x relative to native-language conditions - even as task accuracy remains stable across settings. For socially nuanced classification, English pivots also fail to preserve pragmatic cues, reducing both faithfulness and span agreement. We recommend auditing explanations in the input language, reporting multi-faceted faithfulness metrics beyond lexical overlap, and treating English rationales as communication summaries rather than faithful decision traces.
Somnath Banerjee, Pranav Jha, Rima Hazra +1
Indian Institute of Technology Kharagpur · TCG Crest · National University of Singapore
We propose a new method that allows an LLM to automatically pull in factual knowledge from a knowledge base during token generation. This means that (1)~factual knowledge in the LLM output can be updated without retraining the LLM, (2)~facts in the LLM output can be traced to the knowledge base for transparency and explainability, and (3)~smaller models can achieve the same factual accuracy as larger models. Our core idea is to train the model to produce special tokens that trigger a query to the knowledge base. Our experiments show that our method improves factual grounding in both short and long-form generation, and allows factual revisions to take effect through KB edits rather than parameter updates.
Francois Crespin, Fabian M. Suchanek, Nils Holzenberger
IP Paris, LTCI · Télécom Paris · Institut Polytechnique de Paris, France