Information Retrieval · arXiv:2610.36000 v1 · submitted 26 Oct 2026 · 37 references
RELAY: One Query, Two Roles for Sequential Recommendation
The paper says
“Sequential recommendation must combine semantic preferences with behavioral associations to deliver accurate rankings at low inference cost. Sharing history selection connects preference representation directly to the behavioral evidence used in candidate scoring.”
more of the abstract
We present RELAY, a compact sequential recommender built around one query, two roles: semantic readout and behavioral-memory routing. A single learned query forms a semantic preference representation and reuses its attention to route direct-history and transition evidence into candidate scores. Candidate-conditioned gates calibrate the two memory channels for each candidate. RELAY trains both roles with a single catalog likelihood and produces the final ranking in one forward pass. Full-catalog experiments on six datasets show that RELAY achieves the highest Recall@5 and NDCG@5 among the compared methods on every dataset, improving over the previous best model by 14.22% and 17.88% on average. With 2.044 million predictor parameters, RELAY reduces predictor size by 65.12% and achieves a 1.75 1.79× inference speedup over the previous best model.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 28 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up. RELAY's one-query-two-roles design is a genuine architectural insight that delivers consistent gains with fewer parameters and faster inference; the paper is well-executed and deserves a serious review. This line is wrong
4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 37 references say.
4 minor objections
The paper says, at Abstract and Section 1
Pith says: minor objection 1 of 4
There are minor formatting inconsistencies, such as extra spaces before periods in percentages (e.g., 'by14 .22%and17 .88%on average') and a missing space before 'RELAY' in 'We presentRELAY 1'. These should be corrected.
The paper says, at Table 1
Pith says: minor objection 2 of 4
Baselines are reported as means without standard deviations, while RELAY reports mean ± std from five seeds. Including standard deviations for baselines (where available) would strengthen the comparison, though the consistent margins make this a minor point.
show 2 more
The paper says, at Section 4.4, Table 3
Pith says: minor objection 3 of 4
The quantizer ablation is informative but covers only three of the six datasets. Reporting the full six-dataset results with alternative quantizers (RQ-VAE and RQ-Kmeans) would provide a more complete robustness check, though the three-dataset results are already suggestive.
The paper says, at Appendix B.4
Pith says: minor objection 4 of 4
The inference timing protocol is clearly described, but the paper could add a note about whether GPU memory or batch-size constraints ever become the bottleneck for RELAY given its compact size. This is not needed for publication but would help practitioners.
What would settle it
Re-run the main experiment (full-catalog Recall@5/NDCG@5) on one dataset, e.g., Sports, with a degenerate SID assignment: replace the learned OPQ mapping with a random permutation of the four-code identifiers (preserving the same codebook size and catalog cardinality). If RELAY’s advantage over OneChord and SPRINT shrinks by more than 50% (relative gain halved), then the quantizer quality is indeed load-bearing. If the advantage persists (or shrinks only slightly), the concern about SID dependence is further mitigated.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 28 Oct 2026
- Reader holds up
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee accept: 4 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
RELAY's core idea is that the same attention weights from a single learned query can serve both semantic preference readout and behavioral-memory routing. That is a clean, motivated design, and it works. The paper reports consistent Recall@5 and NDCG@5 improvements over strong SID-based baselines on six domains, with a 65% parameter reduction and roughly 1.75x inference speedup. Those are not cherry-picked results: the gains hold across target popularity and history length, the ablation study is thorough, and the code is provided.
The paper is also careful where it counts. The authors derive the gradient coupling between the two roles (Proposition 3.2), and they work out an exact record-exclusion trick to avoid label leakage from the transition counts (Proposition 3.1). These are not afterthoughts. The experiments include routing interventions (uniform, recency, shuffled) that confirm the learned allocation is driving accuracy.
The main weakness is the reliance on a fixed external quantizer for semantic identifiers, which is a preprocessing step not jointly trained. The paper acknowledges this and partially addresses it by repeating the main comparison with two other quantizers (RQ-VAE, RQ-kmeans); RELAY still leads. That is a reasonable scope condition for a paper focused on the architecture, not the tokenizer. I do not see a load-bearing flaw.
This is a solid, clearly written paper that advances the state of the art in efficient sequential recommendation. It should go to peer review. I would bring it to a reading group and would cite it in related work.
The strongest objection
After thorough review, no internal inconsistency, hidden assumption, or methodological flaw has been found that would undermine the central claim. The reliance on a fixed external quantizer (SID generation) is openly stated and partially defended via a three-dataset quantizer ablation (Table 3). The remaining limitation—that the main results use only OPQ on all six datasets—is a typical scope condition for generative recommendation papers and does not invalidate the reported improvements. The theoretical derivations (record exclusion, gradient coupling) are sound, the code is provided, and the ablation experiments (memory channels, current-record exclusion, routing interventions) consistently support the design choices. The reported accuracy gains, parameter reduction, and speedup are consistent with the architecture and are credible. No load-bearing concern that would change the verdict has been identified.
The reader flagged the fixed external quantizer as the weakest assumption, which is a reasonable observation. However, the paper partially addresses this with a quantizer ablation (Table 3 lower panel) on three datasets, showing RELAY outperforms baselines with OPQ, RQ-VAE, and RQ-Kmeans. The remaining datasets (Scientific, Instruments, Games) are not tested with alternative quantizers, making the generalization claim slightly weaker but not internally inconsistent. No other concern—no hidden overfitting, no mismatched baselines, no theoretical error—was found. The paper’s code, derivations, and ablation breadth (memory channels, exclusion, routing interventions) provide strong support. Therefore, the reader’s ACCEPT verdict is justified; the quantizer concern does not warrant a change.
The referee's report
If the reported results hold, RELAY demonstrates that shared history selection can unify preference representation and behavioral evidence within a compact, fast model, achieving state-of-the-art ranking accuracy at substantially lower parameter count and inference cost. The theoretical derivation of record-excluded catalog learning (Proposition 3.1) and the gradient coupling (Proposition 3.2) provide a principled basis for this design. The consistent gains across six datasets, along with extensive ablations and routing interventions, support the claim that the shared-attention mechanism is effective and robust to quantizer choice. This work advances the efficient generation paradigm for sequential recommendation.
Pith's reading of the paper (a machine's summary, not the authors')
The paper argues that a single learned query can serve both semantic preference reading and behavioral memory routing through shared attention, unifying two traditionally separate tasks in sequential recommendation. This design, called RELAY, uses the same history attention weights to form a user preference representation and to allocate direct-history and transition evidence to candidate scores, with candidate-conditioned gates calibrating each channel. Across six Amazon review datasets, RELAY improves Recall@5 by 14.22% and NDCG@5 by 17.88% on average over the prior best model, while shrinking predictor parameters by 65.12% and speeding inference by 1.75–1.79×. The work suggests that compact, single-pass recommenders can jointly learn what to represent and where to apply behavioral evidence.
What carries the argument. The core mechanism is 'one query, two roles' via shared cross-attention: a single learned query $q$ attends to the encoded history $H$ to produce weights $a_j$ that are reused for two purposes. First, the same weights drive a residual cross-attention block that outputs a semantic readout $z$, which is projected and normalized to yield a semantic score $b_\theta(s)$ for every catalog item. Second, the same weights $a_j$ route each historical item's direct-history indicator $d_j(s)$ and its learned transition evidence $t_j(s;C)$ into memory scores $m_D(s)$ and $m_T(s)$. A pair of candidate-conditioned gates $\gamma_D(s,z), \gamma_T(s,z)$ scales these memory channels before adding them to the semantic baseline, producing the final ranking score $f_\theta(s)$ in one pass.
Core claim. RELAY establishes that history selection can be shared between semantic readout and memory routing in one forward pass. A single learnable query defines attention weights over the encoded history; these weights simultaneously aggregate a semantic preference vector (Role I) and route direct-history and transition evidence to catalog candidates (Role II). Candidate-conditioned gates then adjust the strength of each memory channel per candidate. A full-catalog softmax loss trains both roles end-to-end, and an exact record-exclusion procedure prevents the model from exploiting its own transition counts during training. The resulting system achieves higher accuracy than all sixteen compared methods on every dataset, with far fewer parameters and lower inference cost.
The premise it rests on. The paper assumes that the semantic identifiers (SIDs) produced by a fixed OPQ quantizer (trained independently from the recommender) are sufficiently meaningful for the shared attention to simultaneously capture user preferences and route reliable behavioral evidence; if the quantizer groups items poorly, both roles suffer.
If the paper is right (Pith's inferences, not the authors')
Sequential recommenders can unify preference learning and evidence routing within a single cross-attention module, simplifying architecture design.
The 65% parameter reduction and 1.75× inference speedup make RELAY practical for latency-sensitive deployment on resource-constrained platforms.
The record-exclusion technique for transition counts may generalize to other models that use aggregated statistics as labels or features.
Candidate-conditioned gating provides a principled way to calibrate multiple memory channels adaptively per candidate, potentially extensible to richer behavioral signals.
The reliance on a fixed OPQ quantizer for semantic IDs means that RELAY's performance is partially bounded by quantizer quality; end-to-end learned IDs (e.g., DiGR) could further improve accuracy without architectural change.
The ablations show that transition memory contributes more than direct-history memory, suggesting that cross-item associations are more informative than simple repetition in these domains—a finding that could guide future design of behavioral memory.
The multi-query extension (appendix) hints that multiple shared attentions may better handle diverse user interests, but the paper does not explore the trade-off between query count and latency; a systematic study of this frontier is a natural next step.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Sequential recommendation must combine semantic preferences with behavioral associations to deliver accurate rankings at low inference cost. Sharing history selection connects preference representation directly to the behavioral evidence used in candidate scoring. We present RELAY, a compact sequential recommender built around one query, two roles: semantic readout and behavioral-memory routing. A single learned query forms a semantic preference representation and reuses its attention to route direct-history and transition evidence into candidate scores. Candidate-conditioned gates calibrate the two memory channels for each candidate. RELAY trains both roles with a single catalog likelihood and produces the final ranking in one forward pass. Full-catalog experiments on six datasets show that RELAY achieves the highest Recall@5 and NDCG@5 among the compared methods on every dataset, improving over the previous best model by 14.22% and 17.88% on average. With 2.044 million predictor parameters, RELAY reduces predictor size by 65.12% and achieves a 1.75 1.79× inference speedup over the previous best model.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-28
The authors' main claim, as the classifier picked it out 0.97
“Full-catalog experiments on six datasets show that RELAY achieves the highest Recall@5 and NDCG@5 among the compared methods on every dataset, improving over the previous best model by 14.22% and 17.88% on average.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.06 · human participants 0.03 · controlled comparison 0.76 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.02
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Sequential recommendation must combine semantic preferences with behavioral associations to deliver accurate rankings at low inference cost. | background 0.99 | 0.07 | 0.13 | asserted 1.00 |
| Sharing history selection connects preference representation directly to the behavioral evidence used in candidate scoring. | background 0.43 | 0.15 | 0.08 | asserted 0.99 |
| We present RELAY, a compact sequential recommender built around one query, two roles: semantic readout and behavioral-memory routing. | contribution 0.91 | 0.90 | 0.13 | asserted 0.99 |
| A single learned query forms a semantic preference representation and reuses its attention to route direct-history and transition evidence into candidate scores. | method 0.98 | 0.79 | 0.11 | asserted 1.00 |
| Candidate-conditioned gates calibrate the two memory channels for each candidate. | method 1.00 | 0.73 | 0.03 | asserted 1.00 |
| RELAY trains both roles with a single catalog likelihood and produces the final ranking in one forward pass. | method 0.98 | 0.92 | 0.11 | asserted 1.00 |
| Full-catalog experiments on six datasets show that RELAY achieves the highest Recall@5 and NDCG@5 among the compared methods on every dataset, improving over the previous best model by 14.22% and 17.88% on average. | result 1.00 | 0.96 | 0.58 | asserted 1.00 |
| With 2.044 million predictor parameters, RELAY reduces predictor size by 65.12% and achieves a 1.75~1.79$\times$ inference speedup over the previous best model. | result 1.00 | 0.94 | 0.17 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.96
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-28: 1 request, 6,206 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (5)
Figures are the authors', reproduced from the paper.
References (37)
- [1] Zhuo Cai, Shoujin Wang, Peilin Zhou, Min Xu, Julian McAuley, and Fang Chen on Pith
- [2] Yukuo Cen, Jianwei Zhang, Xu Zou, Chang Zhou, Hongxia Yang, and Jie Tang. 2020. Controllable Multi-Interest Framework for Recommendation. arXiv:2005.09347 [cs.IR] on PithPith's reviewarXiv
- [3] Junchen Fu, Xuri Ge, Alexandros Karatzoglou, Ioannis Arapakis, Suzan Ver- berne, Joemon M. Jose, and Zhaochun Ren. 2026. Differentiable Semantic ID for Generative Recommendation. arXiv:2601.19711 [cs.IR] on PithPith's reviewarXiv
- [4] Zitian Guo, Yupeng Hou, Clark Mingxuan Ju, Neil Shah, and Julian McAuley. 2026. MLPs are Efficient Distilled Generative Recommenders. arXiv:2605.12617 [cs.IR] on PithPith's reviewarXiv
- [5] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk on Pith
- [6] Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. 2022. Learn- ing Vector-Quantized Item Representation for Transferable Sequential Recom- menders. arXiv:2210.12316 [cs.IR] on PithPith's reviewarXiv
- [7] Yupeng Hou, Jiacheng Li, Xiangjun Fu, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. 2024. Bridging Language and Items for Retrieval and Recom- mendation: Benchmarking LLMs as Semantic Encoders. arXiv:2403.03952 [cs.IR] on PithPith's reviewarXiv
- [8] Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating Long Semantic IDs in Parallel for Recommendation. arXiv:2506.05781 [cs.IR] on PithPith's reviewarXiv
- [9] Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. 2022. Towards Universal Sequence Representation Learning for Recom- mender Systems. arXiv:2206.05941 [cs.IR] on PithPith's reviewarXiv
- [10] Yupeng Hou, Jianmo Ni, Zhankui He, Noveen Sachdeva, Wang-Cheng Kang, Ed H. Chi, Julian McAuley, and Derek Zhiyuan Cheng. 2025. ActionPiece: Contextually Tokenizing Action Sequences for Generative Recommendation. arXiv:2502.13581 [cs.IR] on PithPith's reviewarXiv
- [11] Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recom- mendation. arXiv:1808.09781 [cs.IR] on PithPith's reviewarXiv
- [12] Chao Li, Zhiyuan Liu, Mengmeng Wu, Yuchi Xu, Pipei Huang, Huan Zhao, Guo- liang Kang, Qiwei Chen, Wei Li, and Dik Lun Lee. 2019. Multi-Interest Network with Dynamic Routing for Recommendation at Tmall. arXiv:1904.08030 [cs.IR] on PithPith's reviewarXiv
- [13] Zesheng Li, Chengchang Pan, and Honggang Qi. 2026. BRIDGE: Behavior-Guided Residual Integration with Dual-Frequency Graph Evidence. arXiv:2605.22073 [cs.IR] on PithPith's reviewarXiv
- [14] Zesheng Li, Chengchang Pan, and Honggang Qi. 2026. OneChord: Distill What You Rank, Recommend in One Pass. arXiv:2610.25805 [cs.IR] on PithPith's reviewarXiv
- [15] Enze Liu, Bowen Zheng, Cheng Ling, Lantao Hu, Han Li, and Wayne Xin Zhao on Pith
- [16] Shuo Liu, An Zhang, Guoqing Hu, Hong Qian, and Tat-Seng Chua. 2024. Prefer- ence Diffusion for Recommendation. arXiv:2410.13117 [cs.IR] on PithPith's reviewarXiv
- [17] Zhao Liu, Yichen Zhu, Yiqing Yang, Xiao Lv, Guoping Tang, Rui Huang, Qiang Luo, Ruiming Tang, Kun Gai, and Guorui Zhou. 2025. DiffGRM: Diffusion-based Generative Recommendation Model. arXiv:2510.21805 [cs.IR] on PithPith's reviewarXiv
- [18] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel. 2015. Image-based Recommendations on Styles and Substitutes. arXiv:1506.04757 [cs.CV] on PithPith's reviewarXiv
- [19] Liudmila Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Doro- gush, and Andrey Gulin. 2017. CatBoost: Unbiased Boosting with Categorical Features. arXiv:1706.09516 [cs.LG] on PithPith's reviewarXiv
- [20] Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan H. Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheswaran Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. arXiv:2305.05065 [cs.IR] on PithPith's reviewarXiv
show 17 more
- [21] Pengjie Ren, Zhumin Chen, Jing Li, Zhaochun Ren, Jun Ma, and Maarten de Rijke. 2018. RepeatNet: A Repeat Aware Neural Recommendation Machine for Session-Based Recommendation. arXiv:1812.02646 [cs.IR] on PithPith's reviewarXiv
- [22] Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Fac- torizing Personalized Markov Chains for Next-Basket Recommendation. InPro- ceedings of the 19th International Conference on World Wide Web. Association for Computing Machinery, 811–820. doi:10.1145/17... on Pithdoi:10.1145/1772690.1772773
- [23] Ke Shi, Yao Zhang, Feng Guo, Jinyuan Zhang, JunShuo Zhang, Shen Gao, and Shuo Shang. 2026. FAVE: Flow-based Average Velocity Establishment for Se- quential Recommendation. arXiv:2604.04427 [cs.IR] on PithPith's reviewarXiv
- [24] Teng Shi, Chenglei Shen, Weijie Yu, Shen Nie, Chongxuan Li, Xiao Zhang, Ming He, Yan Han, and Jun Xu. 2025. LLaDA-Rec: Discrete Diffusion for Parallel Semantic ID Generation in Generative Recommendation. arXiv:2511.06254 [cs.IR] on PitharXiv
- [25] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang on Pith
- [26] Jiaxi Tang and Ke Wang. 2018. Ranking Distillation: Learning Com- pact Ranking Models With High Performance for Recommender System. arXiv:1809.07428 [cs.LG] on PithPith's reviewarXiv
- [27] Puji Wang, Yingchen Zhang, Ruqing Zhang, Jiafeng Guo, and Xueqi Cheng. 2026. One-Pass Decoding for Generative Recommendation with WFST-Constrained A* Search. InProceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval. Asso... on Pithdoi:10.1145/3805712.3809884
- [28] Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2024. Learnable Item Tokenization for Generative Recommendation. arXiv:2405.07314 [cs.IR] on PitharXiv
- [29] Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan. 2018. Session-Based Recommendation with Graph Neural Networks. arXiv:1811.00855 [cs.IR] on PithPith's reviewarXiv
- [30] Yuhao Yang, Zhi Ji, Zhaopeng Li, Yi Li, Zhonglin Mo, Yue Ding, Kai Chen, Zijian Zhang, Jie Li, Shuanglong Li, and Liu Lin. 2025. Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations. arXiv:2503.02453 [cs.IR] on PithPith's reviewarXiv
- [31] Yuanbo Zhao, Ruochen Liu, Senzhang Wang, Jun Yin, Yuxin Dong, Huan Gong, Hao Chen, Shirui Pan, and Chengqi Zhang. 2026. From Token Gen- eration to Item Ranking: Direct Generative Recommendation with Semantic IDs. arXiv:2602.07847 [cs.IR] on PithPith's reviewarXiv
- [32] Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2023. Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation. arXiv:2311.09049 [cs.IR] on PithPith's reviewarXiv
- [33] Tianyu Zhu, Yansong Shi, Yuan Zhang, Yihong Wu, Fengran Mo, and Jian-Yun Nie. 2023. Collaboration and Transition: Distilling Item Transitions into Multi- Query Self-Attention for Sequential Recommendation. arXiv:2311.01056 [cs.IR] A Derivations and Additional Method Details A.... on PithPith's reviewarXiv
- [2015] Session-based Recommendations with Recurrent Neural Networks. arXiv:1511.06939 [cs.LG] on PithPith's reviewarXiv
- [2019] BERT4Rec: Sequential Recommendation with Bidirectional Encoder Rep- resentations from Transformer. arXiv:1904.06690 [cs.IR] on PithPith's reviewarXiv
- [2024] Generative Recommender with End-to-End Learnable Item Tokenization. arXiv:2409.05546 [cs.IR] on PithPith's reviewarXiv
- [2026] SPRINT: Single-Step Generative Recommendation via Average Probability Velocity. arXiv:2609.34306 [cs.IR] on PithPith's reviewarXiv
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment