Research map / Large language models / Alignment and preferences
Align³GR: Unified Multi-Level Alignment for LLM-based Generative Recommendation
Wencai Ye, Mingjie Sun, Shuhang Chen, Wenjin Wu, Peng Jiang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use dual user-item tokenization instead of item-only (ablation) | Align3GR · Instruments | Recall@10: 0.1322 -> 0.1390 | Improved |
| add collaborative features to item tokenization (ablation) | Align3GR · Instruments | Recall@10: 0.1322 -> 0.1346 | Improved |
| add user-item alignment (U2I loss) to dual tokenization (ablation) | Align3GR · Instruments | Recall@10: 0.1390 -> 0.1426 | Improved |
| add C1-C3 implicit alignment tasks to SEQ (ablation) | Align3GR · Instruments | Recall@10: 0.1329 -> 0.1344 | No clear effect |
| add item alignment task B1 (ablation) | Align3GR · Instruments | Recall@10: 0.1344 -> 0.1399 | Improved |
| inject user SCID into prompts (ablation) | Align3GR · Instruments | Recall@10: 0.1399 -> 0.1417 | Improved |
| add bidirectional user alignment B2 (ablation) | Align3GR · Instruments | Recall@10: 0.1417 -> 0.1442 | Improved |
| add self-play to Softmax-DPO (ablation) | Align3GR · Instruments | Recall@10: 0.1295 -> 0.1356 | Improved |
| add progressive easy-to-hard training to SP-DPO (ablation) | Align3GR · Instruments | Recall@10: 0.1356 -> 0.1396 | Improved |
| add progressive training to RF-DPO (ablation) | Align3GR · Instruments | Recall@10: 0.1414 -> 0.1442 | Improved |
| add real-world feedback DPO after progressive SP-DPO (ablation) | Align3GR · Instruments | Recall@10: 0.1396 -> 0.1442 | Improved |
| deploy Align3GR instead of TIGER online | TIGER · Industrial ad platform | Recall@100 / revenue: 0.229 -> 0.242 Recall@100; revenue +0.555% -> +1.432% | Improved |
Closest papers by meaning
- Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.91)
- ConvRec-R1: Training LLM-based Conversational Recommender Systems with Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · similarity 0.87)
- LLM Alignment as Retriever Optimization: An Information Retrieval Perspective (ICML 2025 · similarity 0.87)
- Mult-DPO: Multinomial Direct Preference Optimization for Recommender Systems (NeurIPS 2026 · similarity 0.87)
- Robust LLM Alignment via Distributionally Robust Direct Preference Optimization (NeurIPS 2025 · similarity 0.87)
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.
