Research map / Large language models / Reasoning and chain of thought
IGD: Token Decisiveness Modeling via Information Gain in LLMs for Personalized 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 IGD, which measures item-token decisiveness via information gain over the item prefix tree and uses it to downweight zero-IG tokens in LLM4Rec fine-tuning and to rebalance beam-search scores, improving BIGRec and D3 on four Amazon datasets.
Problem
LLM-based recommenders treat all item tokens equally during likelihood-based fine-tuning and beam-search decoding, although many tokens carry little item-discriminating information and high-logit low-decisiveness tokens bias both stages.
Approach
Model item generation as a decision process over the item prefix tree, define each token's decisiveness as its information gain (entropy reduction over candidate items), then downweight zero-IG tokens in the tuning loss and rescale log-probabilities by an IG-based factor during beam search.
Key results
- IGD on BIGRec improves HR@10 by an average of 20.9% and NDCG@10 by 21.5% across four Amazon datasets.
- IGD on D3 improves HR@10 by 18.9% and NDCG@10 by 20.1% on average over the D3 backbone, outperforming Pos and CFT token-handling baselines.
- With LLaMA3-8B (QLoRA), D3+IGD improves HR@10 by 26.8% and NDCG@10 by 26.4% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IGD token reweighting to D3 | D3 with Qwen2.5-1.5B · CDs, Games, Toys, Books | HR@10: CDs 0.1040 -> 0.1092; Games 0.0773 -> 0.0946; Toys 0.1029 -> 0.1082; Books 0.0315 -> 0.0424 | Helped |
| Add IGD token reweighting to D3 | D3 with Qwen2.5-1.5B · CDs, Games, Toys, Books | NDCG@10: CDs 0.0767 -> 0.0801; Games 0.0477 -> 0.0598; Toys 0.0698 -> 0.0726; Books 0.0228 -> 0.0313 | Helped |
| Add IGD to D3 | D3 with LLaMA3-8B · CDs, Games, Toys, Books | HR@10: CDs 0.1062 -> 0.1179; Games 0.0832 -> 0.1137; Toys 0.1096 -> 0.1311; Books 0.0299 -> 0.0418 | Helped |
| Add decoding reweighting (ablation) | D3+IGD without decoding reweighting · CDs, Games, Toys, Books | HR@10: CDs 0.1077 -> 0.1092; Games 0.0942 -> 0.0946; Toys 0.1063 -> 0.1082; Books 0.0422 -> 0.0424 | Helped |
| Add tuning reweighting (ablation) | D3+IGD without tuning reweighting · CDs, Games, Toys, Books | HR@10: CDs 0.1041 -> 0.1092; Games 0.0790 -> 0.0946; Toys 0.1060 -> 0.1082; Books 0.0318 -> 0.0424 | Helped |
| Add tuning and decoding reweighting (ablation) | D3 without either IGD component · CDs, Games, Toys, Books | HR@10: CDs 0.1040 -> 0.1092; Games 0.0773 -> 0.0946; Toys 0.1029 -> 0.1082; Books 0.0315 -> 0.0424 | Helped |
| Use binary rather than linear IG token weighting (ablation) | IGD-T · CDs, Games, Toys, Books | HR@10: CDs 0.1072 -> 0.1077; Games 0.0800 -> 0.0942; Toys 0.1034 -> 0.1063; Books 0.0339 -> 0.0422 | Helped |
| Increase decoding decisiveness weight α (ablation) | IGD decoding · CDs, Toys, Games, Books | FWR: CDs 0.364 -> 0.265; Toys 0.588 -> 0.551; Games 0.340 -> 0.296; Books 0.382 -> 0.259 | Helped |
| Increase decoding decisiveness weight α (ablation) | IGD decoding · CDs, Toys, Games, Books | ISE: CDs 2.65 -> 2.92; Toys 2.67 -> 2.95; Games 2.95 -> 3.08; Books 2.75 -> 2.99 | Helped |
Surprising findings
- Over 50% of item tokens (55.96% to 72.97% across datasets) have zero information gain, yet they receive very high logits and dominate tuning and decoding.
- Increasing the decoding strength alpha does not always improve accuracy, indicating a trade-off between sequence likelihood and prioritizing high-decisiveness tokens.
Limitations
- Only IG is used as the decisiveness metric; the IG distribution is highly skewed, making continuous reweighting hard.
- Analysis is restricted to text item tokens and does not cover semantic ID or multimodal tokens.
- Beta and alpha hyperparameters must be searched per dataset.
Experimental setup
Datasets: Amazon CDs Amazon Games Amazon Toys Amazon Books Steam
Benchmarks: Amazon CDs Amazon Games Amazon Toys Amazon Books Steam
Metrics: HR@5 HR@10 NDCG@5 NDCG@10
Baselines: GRU4Rec SASRec LRURec BIGRec D3 Position Normalization Causal Fine-tuning
Builds on: D3 BIGRec Causal Fine-tuning Position Normalization QLoRA
Closest papers by meaning
- Language Ranker: A Lightweight Ranking framework for LLM Decoding (NeurIPS 2025 · similarity 0.87)
- OrthoRank: Token Selection via Sink Token Orthogonality for Efficient LLM inference (ICML 2025 · similarity 0.85)
- Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for RL-based LLM Recommenders (NeurIPS 2026 · similarity 0.85)
- Efficient Inference for Large Language Model-based Generative Recommendation (ICLR 2025 · similarity 0.85)
- Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.85)
- Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.84)
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.