The integration of Large Language Models (LLMs) into explainable recommendation systems often leads to a performance-efficiency trade-off in end-to-end architectures, where joint optimization of ranking and explanation can result in suboptimal compromises. To resolve this, we propose Prism, a novel decoupled framework that rigorously separates the recommendation process into a dedicated ranking stage and an explanation generation stage. This decomposition ensures that each component is optimized for its specific objective, eliminating inherent conflicts in coupled models. Inspired by knowledge distillation, Prism leverages a powerful, instruction-following teacher LLM (FLAN-T5-XXL) as an Oracle to produce high-fidelity explanatory knowledge. A compact, fine-tuned student model (BART-Base), the Prism, then specializes in synthesizing this knowledge into personalized explanations. Our extensive experiments on benchmark datasets reveal a key finding: the distillation process not only transfers knowledge but also acts as a noise filter. Our 140M-parameter Prism model significantly outperforms its 11B-parameter teacher in human evaluations of faithfulness and personalization, demonstrating an emergent ability to correct hallucinations present in the teacher's outputs. While achieving a 24x speedup and a 10x reduction in memory consumption, our analysis validates that decoupling, coupled with targeted distillation, provides an efficient and effective pathway to high-quality, and perhaps more importantly, trustworthy explainable recommendation.
Figures & tables
Figure 1. The overall framework of Prism. The offline stage consists of a teacher phase for data creation via knowledge distillation and a student phase for model fine-tuning. The online stage demonstrates how Prism functions as a decoupled module alongside any SOTA recommender.
Statistic
MovieLens-1M
Yelp
# Users
6,040
1,987,929
# Items
3,883
150,346
# Train Sequences
894,752
1,418,452
# Test Sequences
99,417
157,606
Table 1. Statistics of the processed datasets.
Model
Params
Latency (ms)
Peak GPU (GB)
FLAN-T5-XXL
11B
4612.92
20.60
Prism
140M
190.30
1.91
Improvement
≈ 78 × smaller
≈ 24.2 × faster
≈ 10.8 × lower
Table 2. Efficiency comparison. Latency is the average time to generate one explanation over 100 runs.
Dataset
Model
Automatic Metrics
Human Evaluation
ROUGE-L
GPTScore
BS-F1
Persuasive.
Personal.
Faithful.
Yelp
FLAN-T5-XXL (11B)
0.2761
68.12
0.2679
3.01 ± 0.88
3.07 ± 0.92
2.87 ± 1.01
Att2Seq
0.1653
63.91
0.2377
2.98 ± 0.85
2.76 ± 0.90
2.95 ± 0.98
BART-Base (140M)
0.1607
62.59
0.2599
3.05 ± 0.96
3.13 ± 0.81
3.05 ± 0.95
PEPLER
0.2002
67.24
0.3032
3.42 ± 0.85
3.35 ± 0.77
3.11 ± 0.91
Prism (w/o User)
0.2183
69.21
0.3273
3.85 ± 0.65
3.62 ± 0.80
3.87 ± 0.69
Table 3. Main experimental results on both Yelp and MovieLens-1M datasets. For human evaluation, we report mean score ± standard deviation. IAA using Fleiss’ Kappa was 0.75.
Figure 2. Automatic evaluation results on ROUGE-L, GPTScore, and BS-F1 metrics across Yelp and MovieLens-1M datasets.
Figure 3. Human evaluation results on persuasiveness, personalization, and faithfulness dimensions.
Context
User History: … E.T., Star Wars … , Recommended Item: Back to the Future
A: Golden Explanation
Back to the Future is based on "The Wizard of Oz" and influenced by "The Phantom Menace".
B: FLAN-T5 Prediction
Back to the Future is a sci-fi movie influenced by "The Phantom Menace" and "The Wizard of Oz".
C: A Faithful Explanation
It’s a classic 80s sci-fi adventure, similar to other films in your history.
ROUGE-L (A vs. B)
0.75 (Deceptively High)
ROUGE-L (A vs. C)
0.15 (Unfairly Low)
Table 4. High lexical overlap between two incorrect statements (A vs. B) results in a higher ROUGE-L than between a correct and incorrect statement (A vs. C).
User Context
Simulated Source
Input Item
Generated Explanation (Prism)
Analysis
User 2223 History Patterns: Satire, Dark Comedy, 80s Movies
Oracle (Proxy for SOTA)
Desperately Seeking Susan
“The user has a history of watching movies with satirical themes ."
High Fidelity: Successfully identifies the subtle genre link matching user history.
PopRec (Proxy for Baseline)
American Beauty
“I think you’ll like this movie because it’s a comedy ."
Graceful Degradation: Falls back to a broad genre tag; lacks specific personalization.
Random (Proxy for Noise)
Seven Samurai
“The movie is a comedy about sex and relationships."
Sensitivity Verification: The model fails to connect the irrelevant item to history, resulting in hallucination. Proves Prism does not blindly accept all inputs.
Table 5. Case study of Prism’s Plug-and-Play capability. The model (frozen) generates distinct explanations for the same user when fed with recommendations from different simulated rankers. It accurately explains high-quality inputs while exposing the irrelevance of noise inputs.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
User History
Recommended Item
FLAN-T5-XXL (Teacher)
BART-Base (Zero-Shot)
Prism (Ours)
… E.T., Star Wars, The Thing …
Back to the Future
…based on ”The Wizard of Oz’ and influenced by "The Phantom Menace’.
Back to the Future is a great movie.
Back to the Future is a science fiction film from 1985, fitting your interest in 80s sci-fi.
… Animal House, Caddyshack …
American Pie
…similar to The Shining and The Adventures of Pinocchio.
This is a comedy that suits you
American Pie is a comedy that is similar to the movies you have already seen.
… Ferris Bueller’s Day Off …
Sixteen Candles
The user has a history of watching comedies, and Sixteen Candles is based on teen angst.
Sixteen Candles is a movie.
The user has a history of watching movies about high school. Sixteen Candles is based on teen angst.
Appendix
Table 6. Case study of generated explanations. Our Prism demonstrates higher faithfulness and relevance compared to both the powerful-but-hallucinating teacher model (errors marked in red ) and the naive zero-shot student baseline.