Research map / Large language models / Language, safety and interpretability
LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LLM4RSR, which uses an LLM with automatically optimized profile, plan, and memory (via textual gradient descent) to correct high-uncertainty training instances for robust sequential recommenders, which are then retrained.
Problem
Sequential recommenders are trained on unreliable input-target pairs caused by external distractions, and existing data-driven robust SRs are uncertain about infrequently co-occurring item pairs.
Approach
Uses an LLM as a teacher to correct only the most uncertain training instances (selected by RSR confidence) via an item-selection formulation, with its instruction prompt (profile, plan, item memory) optimized automatically by textual gradient descent and UCB-based selection; corrected data is used to retrain RSRs (selective knowledge distillation).
Key results
- LLM4RSR improves all eight backbones on average by 6.66% (ML-1M), 8.93% (CD), 9.17% (Game), and 8.08% (Kindle) relative to plain training.
- Correcting only the top 5% most uncertain instances gives the best accuracy; correcting more than 15% degrades RSR accuracy.
- Removing the memory module causes the largest drop in item-selection F1 (e.g., ML-1M F1 0.6203 full vs 0.4403 without memory).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Correct uncertain training instances with LLM4RSR | BirDRec · ML-1M | NDCG@10: 0.1913 -> 0.1995 | Helped |
| Correct uncertain training instances with LLM4RSR | BirDRec · CD | NDCG@10: 0.0820 -> 0.0869 | Helped |
| Correct uncertain training instances with LLM4RSR | BirDRec · Game | NDCG@10: 0.0492 -> 0.0517 | Helped |
| Correct uncertain training instances with LLM4RSR | BirDRec · Kindle | NDCG@10: 0.0918 -> 0.0972 | Helped |
| Add a profile module (ablation) | LLM4RSR · ML-1M, CD, Game, Kindle | F1-score: 0.5036 -> 0.6203; 0.5598 -> 0.7496; 0.7229 -> 0.8810; 0.5063 -> 0.6427 | Helped |
| Optimize the profile with TGD rather than keeping it static (ablation) | LLM4RSR · ML-1M, CD, Game, Kindle | F1-score: 0.5426 -> 0.6203; 0.6242 -> 0.7496; 0.7547 -> 0.8810; 0.5141 -> 0.6427 | Helped |
| Add a plan module (ablation) | LLM4RSR · ML-1M, CD, Game, Kindle | F1-score: 0.5029 -> 0.6203; 0.5314 -> 0.7496; 0.7186 -> 0.8810; 0.4621 -> 0.6427 | Helped |
| Optimize the plan with TGD rather than keeping it static (ablation) | LLM4RSR · ML-1M, CD, Game, Kindle | F1-score: 0.5754 -> 0.6203; 0.6896 -> 0.7496; 0.7610 -> 0.8810; 0.5282 -> 0.6427 | Helped |
| Add a memory module (ablation) | LLM4RSR · ML-1M, CD, Game, Kindle | F1-score: 0.4403 -> 0.6203; 0.4419 -> 0.7496; 0.6078 -> 0.8810; 0.4561 -> 0.6427 | Helped |
| Optimize memory with TGD rather than keeping it static (ablation) | LLM4RSR · ML-1M, CD, Game, Kindle | F1-score: 0.5912 -> 0.6203; 0.7234 -> 0.7496; 0.8626 -> 0.8810; 0.5499 -> 0.6427 | Helped |
Surprising findings
- Correcting more than 15% of instances (lower-uncertainty ones) can alter data distribution and hurt RSR accuracy.
- Gains on SRs (9.59% average) are larger than on RSRs (6.81% average), since RSRs already partly mitigate unreliable data.
Limitations
- Correction is expensive relative to pure data-driven methods, so only a small fraction (5%) of instances can be corrected.
- Requires re-training RSRs from scratch on corrected data; unlearning-based updates are left to future work.
- Training-set construction for TGD relies on co-occurrence heuristics assumptions about relevance labels.
Experimental setup
Datasets: ML-1M CD Game Kindle
Benchmarks: ML-1M CD Game Kindle
Metrics: HR@5 HR@10 NDCG@5 NDCG@10 MRR Precision Recall F1-score
Baselines: SASRec LRURec RecFormer PO4ISR BERD+ FMLP-Rec STEAM BirDRec SSDRec
Builds on: textual gradient descent TextGrad UCB BERD BirDRec
Closest papers by meaning
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.93)
- SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · similarity 0.92)
- Customizing Language Models with Instance-wise LoRA for Sequential Recommendation (NeurIPS 2024 · similarity 0.89)
- Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.89)
- LLMEmb: Large Language Model Can Be a Good Embedding Generator for Sequential Recommendation (AAAI 2025 · similarity 0.89)
- Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.