Reviews Review 1 Relevance to CIKM 5: excellent Strengths 1. The authors propose Pctx, a personalized context-aware tokenizer for generative recommendation. 2. The paper is well written and easy to follow and understand. 3. The proposed approach outperforms the baselines on three different dataset with significant improvements. 4. The author share the link to the code to ensure reproducibility. 5. The authors make a detailed ablation study about each component of the proposed approach. Weaknesses 1. I do not appreciate that both the implementation details and the detailed discussion of the results are in Appendix A and D. Since the appendix counts into the ten pages, the authors should discuss the results in a detailed way in Section 3.2. 2. (Minor) Try to avoid using both green and red, which are not distinguishable by color-blind people. Overall evaluation 1: weak accept The authors propose Pctx, a new personalized context-aware tokenizer for generative recommender systems. The paper is easy to follow and the approach is well explained. My only concern is about the discussion of the results and of the implementation details, which should be in the main text and not in the Appendix. Review 2 Relevance to CIKM 5: excellent Strengths 1. (S1) The paper introduces an interesting and novel perspective by questioning the conventional assumption that semantic IDs should be static and item-specific, and instead proposes context-conditioned semantic IDs. 2. (S2) The work is well-positioned within the generative recommendation literature and clearly differentiates itself from existing semantic-ID-based approaches. 3. (S3) The paper provides a solid empirical evaluation of recommendation performance, including comparisons against strong baselines, multiple ablation studies, and several complementary analyses. Weaknesses 1. (W1) The paper's central claim on modeling personalized interpretations is not directly validated, as the evaluation mainly focuses on recommendation accuracy rather than interpretation quality. 2. (W2) The contribution of personalized tokenization itself is difficult to isolate from the effects of strong context representations and several auxiliary components introduced in the framework. 3. (W3) While the paper argues that personalized SIDs capture different user interpretations of the same item, the supporting evidence is largely qualitative. Overall evaluation 1: weak accept The paper proposes Pctx, a personalized context-conditioned tokenization framework for generative recommendation. Instead of assigning a fixed semantic ID to each item, the proposed approach constructs personalized semantic IDs conditioned on user context, aiming to capture different user interpretations of the same item. To address the resulting sparsity and ambiguity issues, the framework further introduces SID merging, data augmentation, and multi-facet decoding mechanisms. Experiments on multiple recommendation benchmarks demonstrate consistent improvements over strong generative recommendation baselines. Overall, the paper explores an interesting direction for semantic-ID-based generative recommendation. The motivation of contextualizing semantic IDs through user-specific interpretations is novel and thought-provoking, and the empirical results demonstrate promising performance improvements across several benchmark datasets. However, there are still several issues that could be further clarified or improved: 1. (D1, related to W1) The central claim of the paper is that personalized semantic IDs capture different user interpretations of the same item. However, the current evaluation primarily focuses on recommendation accuracy, while providing limited direct evidence that the learned semantic IDs indeed correspond to distinct interpretations. Although the case studies are helpful, additional analyses examining the semantic consistency or behavioral differences associated with different personalized IDs would strengthen the paper's core claim. 2. (D2, related to W2) It remains difficult to isolate the contribution of personalized tokenization itself from the effects of other components in the framework. The proposed system relies on several additional mechanisms, including context representations derived from DuoRec, SID merging, augmentation, and multi-facet decoding. While the ablation studies are informative, it is still unclear to what extent the observed gains originate from the personalized tokenization mechanism itself rather than from the auxiliary components introduced to support it. Additional controlled comparisons would help clarify the source of the improvements. 3. (D3, related to W3) The paper argues that personalized semantic IDs capture different user interpretations of the same item, but the evidence is largely limited to a qualitative case study. It remains unclear whether the observed behavior generalizes broadly, as no quantitative analysis is provided to verify that different SIDs consistently correspond to meaningful user-specific facets rather than other latent clustering patterns. Additional quantitative analyses demonstrating that different SIDs correspond to distinct and meaningful user-specific facets would further strengthen the paper's central claim. Review 3 Relevance to CIKM 5: excellent Strengths 1. The paper argues that static semantic IDs impose a generic item similarity structure, which may fail to reflect user-specific understanding. 2. Pctx allows mapping the same item to multiple semantic IDs based on user history, which to some extent addresses the limitations of fixed-item-level segmentation. 3. On Amazon's three product categories, Pctx consistently outperforms traditional sequence recommenders and generative recommendation baseline models. Weaknesses 1. The paper states that context representations originate from training data, but it should be explicitly stated that DuoRec training, context clustering, SID construction, and frequency-based merging should all strictly use the training set. Otherwise, using validation/testing interactions in the segmenter construction will leak future information into the semantic IDs. 2. Pctx uses DuoRec to generate context representations. This means the segmenter already benefits from a separately trained recommendation model. The experiments do not explicitly state how much of the performance improvement comes from personalized segmentation and how much comes from the additional knowledge injected by the auxiliary DuoRec encoder. 3. This method comprises multiple stages: pre-training DuoRec, extracting contextual representations, item-by-item clustering, selecting the number of clusters, fusing contextual and item features, RK-Means quantization, merging duplicate SIDs, merging infrequent SIDs, data augmentation, and multi-faceted decoding. The authors acknowledge this complexity in the limitations section, but the main experimental section lacks quantitative experimental analysis of preprocessing costs, memory overhead, etc. Overall evaluation -1: weak reject See Weaknesses 1~3 Review 4 Relevance to CIKM 5: excellent Strengths S1. The paper presents a simple and effective method for multi-facet generative recsys, as well as various practical pieces for making it work well. S2. The results are strong compared to baselines. Statistically significant with gains of 5-10% or more on the included metrics. S3. The paper includes ablations over key components, analysis including with ensembles, and some visuals that make the method clear to understand. Overall the paper reads well and dense with detail. Weaknesses W1. The clustering seems like a weak fit for this problem, which is probably why steps like merging are needed as well as heuristics for selecting centroid count. That being said, the results still are compelling --- although can wonder if this will lead to problems for other datasets. W2. Multi-facet embeddings are extensively studied but not mentioned. There may be some useful ideas there. One example: https://arxiv.org/abs/2103.15339 W3. The facet prefixes are latent and not directly interpretable. It would elevate the work to have a simple baseline for coming up with prefix space, e.g. running a classifier ahead of time over a fixed vocab. It perhaps would be enough! Although if it does worse, then that provides more justification for pctx. Overall evaluation 1: weak accept This paper extends generative recsys to use multi-facet semantic IDs. The implementation is to use RK-means across all items to get token prefixes representing properties of the facet, and finally to cluster within items to get item-specific tokens (i.e. an item can have multiple semantic IDs, as well as multiple prefixes). The paper includes various pieces that make things work, including merging semantic IDs, re-tokenizing sequences using their contextualized semantic IDs, etc. The results are strong compared to baselines. There are a few methodological weaknesses. Using k-means and RK-means are not well justified --- they have specific biases that may not work well for this data, granted there are some helpful ablations. Additionally, there is minimal reference to related work on multi-facet embeddings which have essentially explored similar idea but outside of generative recsys. Metareview Metareview Title Personalized Context-Conditioned Tokenization for Generative Recommendation Authors Qiyong Zhong, Jiajie Su, Yunshan Ma, Julian McAuley and Yupeng Hou Text Reviewers generally found the paper novel, well written, and empirically strong, with consistent gains over strong baselines and extensive ablations. The main concerns are that the claimed personalized interpretations are only indirectly validated, the effect of the auxiliary DuoRec encoder is not fully isolated, and the multi-stage pipeline is relatively complex. These are meaningful limitations, but they do not undermine the main empirical contribution.