Research map / Large language models / Reasoning and chain of thought
Token-Efficient Long-Term Interest Sketching and Internalized Reasoning for LLM-based Recommendation
Zhihao Ding, Jinming Li, Shuai Mu, Jieming Shi
Read the paperCheck your idea against 137,217 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 a token-efficient method for LLM-based recommendation that sketches long-term user interests and internalizes reasoning, addressing long noisy histories and the cost of CoT at inference.
Problem
LLMs struggle to reason over long, noisy user histories spanning hundreds of items, and truncation discards signals needed for preference reasoning; explicit CoT reasoning also adds inference cost for recommendation.
Approach
Compresses long-term interest into a compact sketch and internalizes CoT-style preference reasoning so that inference does not require generating long explicit reasoning chains (details not available in the abstract).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add long-term interest sketch to recent history (ablation) | answer-only fine-tuned Qwen3-4B · Amazon Movies | MAE: 0.7775 -> 0.7695 | Improved |
| add long-term interest sketch to recent history (ablation) | answer-only fine-tuned Qwen3-4B · Amazon Books | MAE: 0.3547 -> 0.3536 | No clear effect |
| add raw past interactions instead of sketch (ablation) | answer-only fine-tuned Qwen3-4B · Amazon Movies | RMSE: 1.3635 -> 1.3723 | Hurt |
| replace sketch with LLM-generated user profile (ablation) | answer-only fine-tuned Qwen3-4B · Amazon Movies | RMSE: 1.3635 -> 1.3872 | Hurt |
| add GRPO-trained explicit CoT over sketch (ablation) | AO-SFT (+Sketch) · Amazon Movies | MAE: 0.7695 -> 0.7577 | Improved |
| add hidden alignment to internalize CoT (ablation) | GRPO-CoT model · Amazon Books | latency (s/sample): 2.22 -> 0.013 | Improved |
| use hidden alignment instead of CE or KD for internalization (ablation) | GRPO-CoT checkpoint · Amazon Movies | MAE: 0.7995 (KD) -> 0.7603 (HA) | Improved |
| add cross-entropy loss to hidden alignment (ablation) | HA · Amazon Books | MAE: 0.3503 -> 0.3525 | Hurt |
| adapt only KV projections instead of all-linear (ablation) | LoRA hidden alignment · Amazon Books | MAE: 0.3505 -> 0.3503 | No clear effect |
| use brief item text rather than full descriptions for sketch (ablation) | SIREN sketch · Amazon Books | MAE: 0.3536 -> 0.3594 | Hurt |
| use distance-aware regression reward instead of binary exact-match (ablation) | GRPO rating reward · Amazon Movies | MAE: 0.8094 -> 0.7577 | Improved |
| use SIREN versus strongest LLM baseline | Exp3RT · Amazon Books | MAE: 0.4001 -> 0.3510 | Improved |
Closest papers by meaning
- Reinforced Latent Reasoning for LLM-based Recommendation (ICLR 2026 · similarity 0.91)
- LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs (AAAI 2024 · similarity 0.89)
- R$^2$ec: Towards Large Recommender Models with Reasoning (NeurIPS 2025 · similarity 0.89)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.88)
- Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.88)
- CoT4Rec: Revealing User Preferences Through Chain of Thought for Recommender Systems (AAAI 2025 · similarity 0.88)
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.
