TL;DR
This paper introduces Pctx, a personalized context-aware tokenizer for generative recommendation models that improves personalization by incorporating user history into tokenization, leading to significant performance gains.
Contribution
It proposes a novel personalized tokenization method that adapts semantic IDs based on user context, addressing the limitations of static, universal item representations.
Findings
Up to 11.44% improvement in NDCG@10 over baselines
Demonstrates effectiveness of personalized tokenization in recommendation
Validates approach on three public datasets
Abstract
Generative recommendation (GR) models tokenize each action into a few discrete tokens (called semantic IDs) and autoregressively generate the next tokens as predictions, showing advantages such as memory efficiency, scalability, and the potential to unify retrieval and ranking. Despite these benefits, existing tokenization methods are static and non-personalized. They typically derive semantic IDs solely from item features, assuming a universal item similarity that overlooks user-specific perspectives. However, under the autoregressive paradigm, semantic IDs with the same prefixes always receive similar probabilities, so a single fixed mapping implicitly enforces a universal item similarity standard across all users. In practice, the same item may be interpreted differently depending on user intentions and preferences. To address this issue, we propose a personalized context-aware…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The authors rationally modeled the phenomenon that different users respond differently to the same item, and obtained a more adaptive and personalized tokenizer. 2. The authors designed a good strategy to merge redundant semantic IDs. 3. The authors tested Pctx's performance and compared it with previous methods.
1. The Redundant Semantic ID Merging section lacks good mathematical formulas, which hinders understanding. 2. The experiment was only tested on two benchmarks. As far as I know, at least TIGER also provides three public benchmarks: beauty, toys, and sports.
* The paper tackles a less explored direction, personalizing the tokenization stage of generative recommendation instead of static tokenizers * Pctx consistently improves NDCG@10 across three datasets, demonstrating effectiveness * The paper ablates key design components (context clustering, data augmentation, redundancy merging), confirming that each contributes to performance.
* : All experiments use only a single backbone (Sentence-T5-base) for GR modeling, which weakens claims of generality. It is unclear whether the method’s gains hold for other architectures. * In GR, the autoregressive model already conditions on user history, so personalization is inherently modeled at inference. It is unclear why injecting user context into item tokenization provides additional benefit. The reported improvements might arise from increased token diversity or implicit data augmen
1. The manuscript is well-organized and the proposed method is clear to understand. 2. The proposed modules, including contextual embedding, semantic ID merging, and data augmentation are significantly effective. 3. The experiments are sufficient along with detailed analysis and discussion.
1. My main concern lies in the motivation of Pctx to incorporate user history into semantic index generation. Since the recommender takes the user history as input, the hidden state ahead the logits head can capture personalized user preference. Moreover, for an anchor item $v_i$, the hidden states corresponding to two different users $u_m$ and $u_n$ won't be the same, and the proposition that *when generating the next semantic IDs, those with the same prefixes inevitably receive similar probabi
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
