SelfDR: Self-Distillation from Reasoning for LLM-Based Recommendation

1 Introduction

• LLM recommendation approaches include in-context learning and supervised fine-tuning, which can align recommendation representations with semantic representations.The paper presents this alignment as a way to harness LLM capabilities for recommendation.Source: §1 · p. 1
• Figure 1 contrasts multi-step reasoning and reasoning-and-recommendation with SelfDR’s self-distillation from reasoning.The proposed paradigm is intended to enhance effectiveness while maintaining inference efficiency.Source: Figure 1 · p. 1
• Reasoning-based recommendation includes multi-step pipelines and single-step generation of long reasoning traces, but these paradigms impose substantial inference overhead.The paper motivates direct recommendation as a way to reduce inference costs while retaining reasoning’s potential benefit for interpreting recommendation data.Source: §1 · p. 2
• SelfDR trains a reasoner to produce targeted rationales for a teacher, then distills the teacher’s deliberated outputs into a direct-recommendation student.The reasoner and recommenders share the same underlying architecture, and a dynamic weighting mechanism adapts the learning signals.Source: §1 · p. 2
• Unlike prior self-distillation settings, SelfDR distills from the reasoning process, reducing the need for explicit intermediate reasoning outputs while targeting effectiveness and efficiency.The framework is designed to improve recommendation quality without relying on stronger external LLMs.Source: §1 · p. 2
• Experiments on three public datasets evaluate SelfDR’s effectiveness, rationality, and efficiency.The introduction reports these as the framework’s demonstrated strengths.Source: §1 · p. 2

2.1 LLM-Based Recommendation

• LLM-based recommendation adapts language models through prompting, fine-tuning, and hybrid collaborative-filtering architectures, while reasoning methods target subtle preference signals in metadata and user text.The cited work includes multi-step pipelines and single-step reasoning approaches.Source: §2.1 · p. 2
• Reasoning-based recommendation uses multi-step text pipelines or test-time reasoning traces, but these increase latency and resource use in large-scale, time-sensitive systems.The paper contrasts those approaches with direct recommendation in a single pass.Source: §2.1 · p. 2

2.2 Knowledge Distillation

• Knowledge distillation transfers teacher knowledge through predictions, internal representations, or relations, and recommendation systems use it to reduce dimensionality and inference latency.LLM distillation has also transferred semantic reasoning into traditional recommenders or smaller recommendation-oriented LLMs.Source: §2.2 · pp. 2–3
• SelfDR’s framework trains a reasoner with recommendation performance as reward, builds a rationale-augmented teacher, and distills its outputs into a direct-recommendation student using the same base LLM.The student learns reasoning-enhanced ranking without producing explicit rationales.Source: Figure 2 · p. 3
• Prior LLM distillation approaches may rely on powerful external teachers and keep large LLMs in the training loop, adding alignment and computational concerns.The cited discussion notes that a poorly aligned or inaccessible teacher may lead to underperforming distillation.Source: §2.2 · p. 3
• Self-distillation uses a model’s own knowledge as a teaching signal, but distillation from reasoning processes remains largely unexplored in LLM-based recommendation.Prior recommendation examples include conventional-model self-distillation and LLM self-distillation from fine-tuned outputs.Source: §2.2 · p. 3

3.1 Overview

• SelfDR trains a reasoner with downstream recommendation performance as reward, forms a rationale-enhanced teacher, and distills its deliberated outputs into a direct-recommendation student using the same backbone LLM.The student learns reasoning-enhanced ranking behavior without relying on explicit rationales at inference.Source: §3.1 · p. 3

3.2 Task Formulation

• Direct recommendation produces the final recommendation from necessary inputs without intermediate steps, targeting inference efficiency.The task forecasts the next item from the candidate set; candidates may be sampled negatives or top-K items from a base ranker.Source: §3.2.1 · pp. 3–4

3.2.1 Direct Recommendation.

• For direct recommendation, the model predicts the next item from user history and a candidate set, ranking candidates by generated log-probabilities of their identifiers.The candidate set may use randomly sampled negatives or top-K items retrieved by a base ranker; in reranking, it need not include the ground-truth item.Source: §3.2.1 · p. 4

3.2.2 Deliberative Recommendation.

• Deliberative recommendation first generates a rationale from user history and the ground-truth item, then includes that rationale with history and candidates for recommendation.The reason-enhanced stage uses the rationale to capture the model’s understanding of past interactions and likely future item characteristics.Source: §3.2.2 · p. 4

3.3 Learning Reasoning Capability for Teacher Construction

• The teacher’s reasoner is trained with GRPO [40] using downstream recommendation accuracy as reward, and its generated rationale is incorporated into the teacher input.The method avoids relying on user reviews, which may be unavailable or irrelevant, and masks rationale phrases overlapping with item titles before reward computation.Source: §3.3 · p. 4

3.4 Self-Distillation for Direct Recommendation

• The student learns from teacher logits rather than generated labels alone, using reverse KL divergence in offline distillation while keeping the teacher fixed.The teacher receives the generated rationale, whereas the direct-recommendation student does not; concise outputs keep training cost modest relative to long reasoning generations.Source: §3.4 · p. 5
• Dynamic weighting emphasizes teacher guidance when informative and reduces it when the teacher ranks the positive item poorly or underperforms the student.Soft gates account for the teacher’s rank relative to a threshold and the teacher–student rank difference, with a penalty for underperforming-teacher cases.Source: §3.4 · p. 5

4 Experiments

• Experiments address baseline performance, reasoner comparisons, self-distillation’s contribution, and training cost and inference efficiency.These are the four research questions posed for the study.Source: §4 · p. 5

4.1.1 Datasets.

• The datasets are Amazon Clothing Shoes and Jewelry, Amazon Home and Kitchen, and MovieLens-1M, processed with 5-core filtering and leave-one-out splitting.The Amazon datasets include product metadata and reviews; ML1M contains movie ratings, user information, and movie metadata.Source: §4.1.1 · p. 5

4.1.2 Evaluation.

• Evaluation uses base-ranker-generated candidates and measures both single-item generation with HitRate@1 and ranked lists with HitRate and NDCG.The positive item is not guaranteed to appear in the candidate set; ranked lists sort candidates by generated identifier-token log-probability.Source: §4.1.2 · p. 6

4.1.3 Baselines.

• Baselines include ID-based and content-based rerankers, direct-recommendation LLMs, multi-step reasoning methods, and reasoning-and-recommendation methods.The listed LLM-based examples include ZSRanker and SOFT, EXP3RT and COT4Rec, and RecSAVER and RecR1.Source: §4.1.3 · p. 6

4.1.4 Implementation Details.

• LLM-based recommenders use LLaMA-3.1-8B-Instruct, and reported results are averaged over three seeds.SASRec generates the top-20 candidates; distillation divides the training set chronologically, uses up to three epochs with early stopping, and sets the threshold k to 5.Source: §4.1.4 · p. 7

4.2 Main Results (RQ1)

• SelfDR significantly outperforms traditional and LLM-based recommenders across all metrics on all three datasets.The authors describe this as evidence of effectiveness and robustness for self-distillation from reasoning within a unified model architecture.Source: §4.2.1 · p. 7

4.2.1 Recommendation List Quality.

• Traditional rerankers often improve on the base ranker, while ID-based models can improve top-ranked accuracy and content-based models can strengthen ranking quality across cutoffs.The authors note that traditional models use textual signals more narrowly than LLMs, which have richer commonsense and world knowledge.Source: §4.2.1 · p. 7
• EXP3RT improves over the base results only at @1 and degrades at larger cutoffs, while COT4Rec remains below SelfDR.The authors attribute EXP3RT’s list-quality issue to often uniformly high predicted ratings and describe COT4Rec’s external-LLM-generated rationales as potentially noisy and not always recommendation-targeted.Source: §4.2.1 · p. 7

4.2.2 Top-1 Generation Accuracy.

• SelfDR consistently achieves the best performance under ranking-based and generation-based evaluation, while RecR1 shows strong top-1 generation accuracy on Clothing and Home but is less effective at fine-grained ranking.Source: §4.2.2 · p. 7

4.3.1 Post-Distillation Performance on Test Set.

• The trained Reasoner achieves the best post-distillation Hg@1 performance across all three datasets among the compared reasoners.The figure compares three external LLMs, untrained LLaMA, and the trained LLaMA.Source: Figure 3 · p. 7, §4.3.1 · p. 7
• A more powerful general-purpose LLM does not necessarily produce better post-distillation recommendations, and external-model rankings vary by dataset.The authors suggest domain specialization or a cognitive gap may make some large-model rationales less learnable by the smaller student.Source: §4.3.1 · p. 8

4.3.2 Teacher Performance on Training Set.

• On 2,000 sampled Clothing training instances, all rationale-enhanced teachers outperform the student before distillation, and the trained Reasoner performs best among the compared reasoners.GPT-5-Chat outperforms GPT-4o-Mini but is not the most effective external Reasoner.Source: §4.3.2 · p. 8, Table 4 · p. 8

4.3.3 Qualitative Comparison of Reasoner-Generated Rationales.

• For the Clothing example, the trained Reasoner produces the most targeted and informative rationale, emphasizing functionality and gift-related intent relevant to the next item.The off-the-shelf LLaMA captures user-history characteristics but misses target-relevant emphasis; Claude connects history to the next item but misses the gift signal.Source: Table 5 · p. 8, §4.3.3 · p. 8

4.4 Effect of Self-Distillation (RQ3)

• The ablation compares instruction tuning alone, supervised-only and distillation-only self-distillation, and fixed-α weighting, with self-distillation epochs held constant.Source: §4.4.1 · p. 8

4.4.1 Ablation on Self-Distillation.

• Dynamic weighting combines informative teacher signals with ground-truth supervision, improving overall performance while avoiding the underperformance of a naïve mixture.Supervised-only training tends to overfit, while distillation-only omits label supervision that stabilizes learning.Source: §4.4.1 · p. 8, Table 6 · p. 9

4.4.2 Impact of Self-Distillation on Output Confidence.

• SelfDR raises confidence in ground-truth items ranked within the top two positions on both training and test sets.The pattern is consistent across the two sets.Source: §4.4.2 · p. 9, Figure 4 · p. 9

• On Clothing, SelfDR achieves the most efficient inference, tied with SOFT, by producing only a single-character identifier; all other baselines are substantially slower in inference.Table 7 reports training time in hours and inference time in ms/instance.Source: §4.5 · p. 9, Table 7 · p. 9
• Despite its two-stage pipeline, SelfDR’s total training time is comparable to prior methods and notably lower than RecR1’s substantially longer GRPO procedure.SelfDR’s reasoner converges quickly, and its one-token recommender output keeps instruction tuning and self-distillation lightweight.Source: §4.5 · p. 9, Table 7 · p. 9
• SelfDR avoids external LLM supervision costs, unlike EXP3RT, COT4Rec, and RecSAVER, which query state-of-the-art LLMs.Source: §4.5 · p. 9

5.1 Hyperparameter Sensitivity Analysis

• On Clothing, α_base = 0.9 performs best; β ≈ 0.1 usually works well, and temperature coefficients typically perform best in [1.0, 2.0].The optimal α_base varies across datasets, while the model is generally robust to temperature choices.Source: §5.1 · pp. 9–10, Figure 5 · p. 10

5.2 Candidate Generator Robustness

• SelfDR achieves the best performance across all metrics on Clothing with LightGCN and ComiRec candidate generators.All three compared methods improve over their original base rankers.Source: §5.2 · p. 10, Table 8 · p. 10

5.3 Out-of-Domain Generalization

• Self-distilled models consistently outperform their non-distilled counterparts when transferred from Clothing to Home and from Home to Clothing.The results support gains beyond the training domain.Source: §5.3 · p. 10, Table 9 · p. 10

• SelfDR trains a reasoner to generate targeted rationales, then uses a same-architecture student to learn from the teacher’s deliberated outputs through dynamic loss.Source: §6 · p. 10
• Experiments on three public datasets show the best recommendation effectiveness while maintaining high inference efficiency.Source: §6 · p. 10
• The study centers on LLaMA-8B and identifies evaluation on different LLM types and scales as future work.The authors describe this as an early exploration of distilling from LLM reasoning for self-evolution.Source: §6 · p. 10

GenAI Usage Disclosure

• Generative AI tools were used only for language editing and polishing of author-written text.The edits addressed grammar, clarity, and fluency.Source: GenAI Usage Disclosure · p. 11
• The tools were not used to generate research ideas, methods, data, experimental results, citations, or scientific claims.Source: GenAI Usage Disclosure · p. 11
• The authors reviewed, verified, and revised all AI-assisted edits and take full responsibility for the work's accuracy, originality, and integrity.Source: GenAI Usage Disclosure · p. 11