Pith.

Artificial Intelligence · arXiv:2610.26035 v1 · submitted 19 Oct 2026 · 44 references

Decision4Rec: A Jev-Inspired Framework of Decision Models on Knowledge and Active Graphs for Multi-Agent Recommender Systems

Marco Polignano, Het Darshan Mehta, Pasquale Lops, Ernesto William De Luca, Giovanni Semeraro

The paper says

“Decision models, such as the commercial Jev, are language models that return one probability for each option of a question instead of generating text. We present Decision4Rec, an open and modular framework that makes decision models useful for recommendation.”

more of the abstract

A knowledge graph writes the logs of each recommendation scenario as short statements on a defined scale, such as "sold: top 10%". A pool of agents, proposes candidates, and a residual ranker keeps the best 30. A decision model, trained once on real choices of three scenarios, reads them through one interface of typed questions, and a fusion step combines its answers with the score of the ranker. Any model with this interface can be used, and we run the commercial Jev, the released open Kev and Kev models trained by us on recommendation requests. The framework with a trained 4B decision model is 25 to 36% above the best trained baseline in NDCG@10 in three sparse Amazon catalogues, one of them never used to train the decision model, and almost tie with baselines on the dense Steam dataset. Most of the gain comes from the pool, and with a trained recommender it comes from what the agents know about its candidates rather than from new candidates. The 27B decision model adds to the pool 5 to 6% above the 0.8B model in three of four domains. A trained 0.8B open model beats the released Jev in eleven of fifteen settings and is never worse. On MovieLens-1M, a dense domain never seen in training, the gain is 5 to 7%. Code, configurations and trained models are available at https://github.com/marcopoli/decision4rec and https://huggingface.co/m-polignano.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 20 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

The passes disagreed. Reader: holds up. Skeptic: holds up only with conditions. Referee: major revision.

Solid framework, provisional headline numbers—worth engaging, but the reader overstates confidence. This line is wrong

2 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 44 references say.

The paper says, at §6.3, §7.1, §7.2, §9.2

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.81):

“These results are exploratory, since several versions were tried on the same users. 7.4 RQ4: trust and cost The probabilities are calibrated but flat(Table 13).”

Pith says: Objection 1 of 2 · major

The central numerical claims (25–36% NDCG@10 improvements on sparse Amazon catalogues, the comparison with Jev, size-scaling results) are all measured on the 2,000 development test users per domain. Section 6.3 states that design choices (fusion weights, λ, feature sets, agent selection) were made using these same users, and that 10,000 fresh users per domain were reserved with an analysis plan written before scoring them.

more of objection 1

Section 9.2 item 1 lists this confirmation as ongoing work. Until those fresh-user results are reported, every headline improvement is provisional. The risk of overfitting to the development split is real: seed spread of the 0.8B model changed fused NDCG@10 by up to 3% (§7.1), and no multiple-comparison correction was applied to the development tables. This is the single most load-bearing unresolved issue. The authors acknowledge it openly, but presenting development-only results as the paper's primary quantitative evidence is premature. The paper should either report the confirmation results or, failing that, state clearly that all findings are preliminary and update the abstract and conclusions to remove definitive claims.

The paper says, at §7.1, Table 8

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

The 27B decision model is trained on only 8,000 records (a random quarter of the 34,667 records used for the 0.8B and 4B models), yet its fused gains over the pool are reported as 4–8% and compared directly with the 0.8B and 4B models trained on all records.

more of objection 2

Table 8 provides the fairer comparison on the same 8,000 records, but the main text (§7.1, the answer to RQ1) first reports the 27B results without consistently noting the reduced training data. The seed spread for the 27B model was not measured, so the significance of its advantage over the 4B model is unclear. The paper should either seed-train the 27B model or clearly state that the comparison is provisional.

5 minor objections

The paper says, at §6.3

Pith says: minor objection 1 of 5

Typographical error: 'TThe lists of all users' should be 'The lists of all users'.

The paper says, at Fig. 7 caption

Pith says: minor objection 2 of 5

The caption reads 'Values from Tables 11 and 10,' but Fig. 7 illustrates the additive gain of the decision model over the pool. Table 11 does contain those values, but Table 10 (MovieLens-1M) also contributes to the figure; the reference should be more precise.

show 3 more

The paper says, at §8, Table 15

Pith says: minor objection 3 of 5

The phrase 'Kev-27B, released' appears in Table 15 but the paper previously uses 'Kev-27B' for the trained version; clarify the notation in the table footnotee or column headings.

The paper says, at §9.1, Limitations

Pith says: minor objection 4 of 5

The authors state that 'the decision model overlaps with the ranker' and give explainability results (AM8). It would be helpful to move the explanation analysis into the main body or at least reference it more prominently, as it directly supports the claim about the ranker explaining most of the model's order.

The paper says, at References

Pith says: minor objection 5 of 5

Reference [20] (Kev contributors) is a GitHub repository, not a peer-reviewed publication; this is acceptable for a technical report, but the paper should note when a reference is a codebase rather than a published article.

What would settle it

Execute the pre-registered confirmation experiment described in Section 9.2, item 1: on the reserved 10,000 fresh users per domain (Video Games, Musical Instruments, Steam, Industrial and Scientific), with a split on global time and with SASRecF re-trained to include also the targets that train the ranker and the decision model. Compute NDCG@10 for Decision4Rec-Full (4B) and for the best trained baseline (SASRecF on Amazon domains; SASRec on Steam). Apply the Holm correction over the nine hypotheses of RQ1 and RQ2. If the three sparse Amazon domains still show a statistically significant improvement (corrected p < 0.05) of at least 20% over the baseline, the central concern is resolved. Report the point estimates and intervals.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 20 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic disagrees: says it holds up only with conditions 34 s · 35,534 tokens read, 3,259 written Central claim rests on development-only results; confirmation on 10k fresh users per domain with global-time split is planned but not yet reported, leaving the headline improvements unvalidated.
  3. Editor wrote the verdict above 10 s · 36,417 tokens read, 840 written
  4. Referee major revision: 2 major objections, 5 minor 26 s · 35,910 tokens read, 2,497 written

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

This is well-executed work, and the reader's description is accurate in substance, but I want to flag a mismatch between the reader's HIGH confidence and the paper's own stated limitations. The headline improvements (25–36% NDCG@10 lift on sparse Amazon catalogues) come from 2,000 development users per domain. The authors themselves acknowledge this clearly (Section 9.2, item 1), and they have a pre-registered plan to confirm on 10,000 fresh users per domain with a global-time split. That confirmation is explicitly listed as ongoing work. The stress-test note is correct: until those numbers are reported, the central quantitative claim is provisional. The paper is also upfront about the global-timeline issue, seed spread, and overlap between the decision model and the ranker. So this is not a paper with hidden flaws—it is a paper that honestly reports results on its development split and says it will do more. The reader's HIGH confidence is premature because the key numbers are not yet independently validated on held-out users. However, the framework itself is novel and well-engineered: a modular pipeline with a knowledge graph that writes scenario-specific behavior statements, a multi-agent pool with a residual ranker, and a trained decision model that reads typed questions. The evaluation is rigorous on the development set—full-catalogue metrics, significance tests, contamination checks, open code and data. The finding that training a small open model beats a large commercial API is robust and practically important regardless of the absolute lift numbers. The paper's own limitations are thorough. My recommendation: this paper should go to peer review. It is serious, honest, and the release quality is high. The referee should ask the authors to report the confirmatory results on fresh users with the global-time split before final acceptance. The reviewer will not need to find new problems; the authors have already done that work. The paper will be a solid contribution once the confirmation is in, and the framework itself is already useful.

The strongest objection

The paper's headline improvement (25–36% NDCG@10 lift on three sparse Amazon catalogues) is measured on 2,000 test users per domain that were part of the development set (design choices including fusion weights, λ, feature sets, and agent selection were made using these same users). The authors reserved 10,000 fresh users per domain and registered an analysis plan 'written before scoring them' (Section 6.3), but those confirmation experiments are explicitly listed as ongoing work (Section 9.2, item 1). Until those results are reported, every claim about the magnitude and even the existence of the improvement is provisional. The risk of overfitting to the development split is real: the paper itself acknowledges that seed spread of the 0.8B model changed fused NDCG@10 by up to 3% and that multiple comparisons were not corrected in the development tables. Moreover, the global-timeline issue (Section 9.1) means absolute values may shift under a proper chronological split, though relative comparisons should be fair. The combination of unconfirmed numbers and acknowledged design-sensitivity makes the central quantitative claim the single most load-bearing unresolved concern. This does not impugn the framework's novelty or the released code, but the reader's verdict of ACCEPT with HIGH confidence is premature given that one of their own planned validation steps remains unexecuted.

The reader's weakest assumption focuses on cross-domain transfer of behavior statements. While that is a legitimate design concern, the paper provides strong evidence that raw counts (which do not rely on percentile cutoffs) transfer equally well (Section 4.2), and the held-out domain results (Industrial and Scientific, MovieLens-1M) already demonstrate that the decision model works without retraining on the target domain. Thus the transfer assumption is empirically supported, not the weakest link. Instead, the most pressing unresolved issue is the lack of confirmation on the reserved 10,000 fresh users. The paper openly states that the confirmation is ongoing, but the headline numbers—which are the basis for the ACCEPT verdict—are from the development split. Given that design choices (fusion weights, λ, feature selection, agent set) were tuned on the same development users, overfitting cannot be ruled out. The paper also reports that seed spread of the 0.8B model caused fused NDCG@10 variation of up to 3%, and that multiple testing corrections were deferred to the confirmation. These are standard good practices, but they mean the current results are preliminary. An ACCEPT verdict with HIGH confidence requires that the confirmation be completed or that the claim be qualified accordingly. The CONDITIONAL recommendation reflects that the framework is promising and well engineered, but the central quantitative claim must pass confirmation before full acceptance. The paper's own honesty about the limitation (Section 9.2) makes this a perfect candidate for a conditional accept: everything else is in order, but the key numbers need independent verification on fresh users.

The referee's report

If the reported gains hold under the planned confirmation on fresh users with a global-time split, the paper offers three important advances: (i) it shows that a decision model trained on real choices from a few source domains can transfer to unseen domains and provide substantial lifts in sparse catalogues; (ii) it demonstrates that most of the improvement can be obtained from a pool of simple knowledge-graph agents and a small residual ranker, making the approach cheap; (iii) it releases the full code, trained models, and configuration, enabling reproducibility and extension. The modular architecture and the careful diagnostic experiments (G1–G7) are methodologically valuable. However, the significance is tempered by the fact that the headline quantitative claims rest entirely on development users, with the planned confirmation study (Section 9.2 item 1) still listed as ongoing work.

Pith's reading of the paper (a machine's summary, not the authors')

Decision4Rec is a modular framework that makes decision models (language models that return choice probabilities instead of generating text) useful for recommendation. The key insight is that a released decision model adds nothing to a trained recommender until it reads statements about how items behave in the scenario—written by a knowledge graph on scenario-specific scales (e.g., 'sold: top 10%')—and is trained on real user choices. With a trained 4B decision model, the framework improves NDCG@10 by 25–36% over the best trained baseline in three sparse Amazon catalogues, one of which was never seen during training. Most of the gain comes from a multi-agent pool that keeps the trained recommender's list and corrects its order with a small residual ranker; the decision model adds further lift at 4B and 27B parameters. A small 0.8B trained model beats the commercial Jev in 11 of 15 settings and is never worse. The authors show that training on real choices matters far more than model size or price, and they release the framework, code, and trained models openly.

What carries the argument. The central object is a trained decision model, specifically Kev (an open implementation of the Jev protocol), fine-tuned on real next-item choices via low-rank adapters and a choice head. It receives a state (user history), a set of up to 30 candidates with short text, and for each candidate five families of behavioural statements written by a never-trained knowledge graph: overall popularity, recent popularity, follow-up, co-purchase, and text similarity, each expressed as position on scenario-specific percentile scales (e.g., 'top 10%', 'often'). The decision model returns a probability per candidate in a single deterministic forward pass. The probabilities are then combined across two random orders via a Luce model to produce a worth for each item, and fused with the score of a residual ranker that corrects the order of a multi-agent pool. The pool always holds the trained recommender's list whole, and other agents (follow-up, co-purchase, text, popularity, etc.) supply candidate features rather than new candidates. This design meets the paper's seven identified requirements: the model reads scenario behaviour, statements are scale-invariant, candidate list integrity is preserved, training records are free of cues, results are reproducible with open weights, and the model is an exchangeable slot.

Core claim. A decision model that is trained on real next-item choices from three recommendation scenarios, and that reads short behavioural statements about each candidate (positioned on scenario-specific scales), can be reused unchanged in a fourth scenario and deliver significant gains over the best trained baseline in sparse catalogues. The framework, called Decision4Rec, separates the task into four modules: a knowledge graph writes logs as statements on scales (e.g., 'overall popularity: top 10%'), a multi-agent pool collects candidates while preserving the trained recommender's list, a residual ranker corrects the order using agent features, and a trained decision model (Kev-4B) re-ranks the best 30 candidates. On two Amazon source domains (Video Games, Musical Instruments) the fused score beats SASRecF by 25% and 36% respectively; on Industrial & Scientific, a held-out sparse domain, it beats the same baseline by 31% even though the decision model was never trained there. The pool alone accounts for most of the gain (17–26% above SASRecF), and the decision model adds 5–8% on top at 4B and 5–6% more at 27B. Training a small 0.8B model beats the commercial Jev in eleven of fifteen settings, showing that training on real choices is the decisive factor.

The premise it rests on. The framework's ability to transfer a decision model to a new domain relies on the assumption that behavioural statements expressed as positions on scenario-specific scales (e.g., 'top 10%') carry a stable meaning across domains—the paper tests raw counts as a control and finds no significant difference, so the scale itself is not proven to be the cause of transfer, but the whole reuse mechanism collapses if the statements fail to represent item behaviour consistently in an unseen domain.

If the paper is right (Pith's inferences, not the authors')

A trained open decision model can replace a commercial API (Jev) for recommendation, with better accuracy and lower cost, enabling reproducible and deployable systems.

In sparse catalogues (few interactions per item), the Decision4Rec framework with a pool and a trained decision model offers a clear accuracy improvement over the strongest sequential models (SASRec, GRU4Rec), suggesting that behavioural statements from logs provide complementary signal.

The multi-agent pool, which needs no language model at runtime, supplies most of the gain and is cheap to deploy; the decision model can then be upgraded independently when a better open model becomes available.

Because the decision model is trained once on three source domains and transfers to unseen sparse domains, a single training effort can serve many related scenarios without per-domain retraining.

The paper's finding that the decision model overlaps heavily with the residual ranker (80–89% of score variance explained by ranker features) suggests that the current behavioural statements and agent features already encode most of the signal the small decision model learns; larger models (27B) break away slightly. Pushing beyond this plateau likely requires richer item descriptions, user profile

The framework's modularity means that failure modes are isolable: if a new domain has no log data (catalogue only), the knowledge graph provides text similarity and the decision model's gain is modest (27–34% above content kNN). The paper's strongest result requires logs to build a pool. A practitioner should therefore invest in gathering minimal interaction data before deploying the decision mode

The calibration analysis shows that probabilities are flat (the next item is rarely singled out). This implies that using conformal prediction to produce a shortlist would require many candidates, limiting practical use for explicit risk-aware recommendations unless the model's confidence sharpens with size or training data.

Because the paper holds out the last item per user (not global time), the absolute NDCG values may shift under temporal splitting, but the relative ordering of methods is likely robust since all methods share the split. The ongoing confirmation on fresh users with a temporal split will clarify this.

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

Every value above, in the review's public record (JSON)

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.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Decision models, such as the commercial Jev, are language models that return one probability for each option of a question instead of generating text. We present Decision4Rec, an open and modular framework that makes decision models useful for recommendation. A knowledge graph writes the logs of each recommendation scenario as short statements on a defined scale, such as "sold: top 10%". A pool of agents, proposes candidates, and a residual ranker keeps the best 30. A decision model, trained once on real choices of three scenarios, reads them through one interface of typed questions, and a fusion step combines its answers with the score of the ranker. Any model with this interface can be used, and we run the commercial Jev, the released open Kev and Kev models trained by us on recommendation requests. The framework with a trained 4B decision model is 25 to 36% above the best trained baseline in NDCG@10 in three sparse Amazon catalogues, one of them never used to train the decision model, and almost tie with baselines on the dense Steam dataset. Most of the gain comes from the pool, and with a trained recommender it comes from what the agents know about its candidates rather than from new candidates. The 27B decision model adds to the pool 5 to 6% above the 0.8B model in three of four domains. A trained 0.8B open model beats the released Jev in eleven of fifteen settings and is never worse. On MovieLens-1M, a dense domain never seen in training, the gain is 5 to 7%. Code, configurations and trained models are available at https://github.com/marcopoli/decision4rec and https://huggingface.co/m-polignano.

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 · 2026-10-20

The authors' main claim, as the classifier picked it out 0.97

“The framework with a trained 4B decision model is 25 to 36% above the best trained baseline in NDCG@10 in three sparse Amazon catalogues, one of them never used to train the decision model, and almost tie with baselines on the dense Steam dataset.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.46 · human participants 0.07 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.99

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.

SentenceRoleThe authors' own workStates the main conclusionStrength
Decision models, such as the commercial Jev, are language models that return one probability for each option of a question instead of generating text. background 1.00 0.05 0.02 asserted 1.00
We present Decision4Rec, an open and modular framework that makes decision models useful for recommendation. contribution 1.00 0.84 0.08 asserted 1.00
A knowledge graph writes the logs of each recommendation scenario as short statements on a defined scale, such as "sold: top 10%". method 0.98 0.34 0.02 asserted 0.99
A pool of agents, proposes candidates, and a residual ranker keeps the best 30. method 0.99 0.63 0.03 asserted 0.94
A decision model, trained once on real choices of three scenarios, reads them through one interface of typed questions, and a fusion step combines its answers with the score of the ranker. method 1.00 0.76 0.05 asserted 0.98
Any model with this interface can be used, and we run the commercial Jev, the released open Kev and Kev models trained by us on recommendation requests. method 0.94 0.82 0.03 asserted 0.96
The framework with a trained 4B decision model is 25 to 36% above the best trained baseline in NDCG@10 in three sparse Amazon catalogues, one of them never used to train the decision model, and almost tie with baselines on the dense Steam dataset. result 1.00 0.95 0.56 asserted 1.00
Most of the gain comes from the pool, and with a trained recommender it comes from what the agents know about its candidates rather than from new candidates. interpretation 0.90 0.93 0.25 asserted 1.00
The 27B decision model adds to the pool 5 to 6% above the 0.8B model in three of four domains. result 1.00 0.92 0.07 asserted 1.00
A trained 0.8B open model beats the released Jev in eleven of fifteen settings and is never worse. result 1.00 0.92 0.13 asserted 1.00
On MovieLens-1M, a dense domain never seen in training, the gain is 5 to 7%. result 1.00 0.94 0.11 asserted 1.00
Code, configurations and trained models are available at https://github.com/marcopoli/decision4rec and https://huggingface.co/m-polignano. contribution 0.93 0.46 0.02 asserted 0.99

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.92

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, 2026-10-20: 1 request, 8,963 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 (7)
Figure 1 of the paper
Figure 1. Two ways to use a language model for ranking. A generative model writes a list as
Figure 2 of the paper
Figure 2. Example of a choice question and its answer. The probabilities are invented for illustration and are not measured. processed once, many questions are answered in one pass, and asking a question alone or together with two others changed its answer by at most 0.005 in our tests. Property (iii) makes it cheap to ask the same choice twice in different orders (Section 5.6). What we do not assume. The architecture and training of Jev are not public, and its calibration is reported as good for choices but not for yes or no questions or fine scales [3, 6]. We therefore measure calibration…
Figure 3 of the paper
Figure 3. The same raw number means different things in different scenarios. Both bars use the
Figure 4 of the paper
Figure 4. Architecture of Decision4Rec: four modules with fixed contracts (in italics), designed for the ActiveGraph runtime and run as a batch pipeline in the experiments. (1) The knowledge graph of the scenario, never trained, writes statements on scales of the scenario. (2) The multi￾agent pool, with the trained recommender as one agent; a residual ranker corrects its score and keeps 30 candidates. (3) The decision model, a slot that any model with the typed-question interface can fill. (4) Fusion with the ranker score. 5.2 The knowledge graph of the scenario, read as statements This…
Figure 5 of the paper
Figure 5. What each part of the system may see, for one user with history
Figure 6 of the paper
Figure 6. The runtime. The boxes in the first row are the main events in the log, and the
Figure 7 of the paper
Figure 7. What the trained decision model adds to the pool, by size. Each bar is the change

Figures are the authors', reproduced from the paper.

References (44)
  1. [1] Ethan Bito, Yongli Ren, and Estrid He. One pass, any order: Position-invariant listwise reranking for LLM-based recommendation. InProceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2026 on Pith
  2. [2] Gordon V. Cormack, Charles L. A. Clarke, and Stefan Büttcher. Reciprocal rank fusion outperforms Condorcet and individual rank learning methods. InProceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, 2009 on Pith
  3. [3] Tobias Deußer, Lorenz Sparrenberg, and Rafet Sifa. Evaluating and benchmarking the System One model Jev.arXiv preprint arXiv:2609.37647, 2026 on PithPith's reviewarXiv
  4. [4] Dario Di Palma, Felice Antonio Merra, Maurizio Sfilio, Vito Walter Anelli, Fedelucio Narducci, and Tommaso Di Noia. Do LLMs memorize recommendation datasets? a preliminary study on MovieLens-1M.arXiv preprint arXiv:2505.10212, 2025 on PithPith's reviewarXiv
  5. [5] Feiyu Duan, Jiayu Lin, Jia Wang, Jun Xiang, Jialiang Wu, Xinnong Zhang, Hanqi Yan, Siyuan Wang, and Zhongyu Wei. General decision models: Benchmarking and insights beyond Jev.arXiv preprint arXiv:2610.03935, 2026 on PithPith's reviewarXiv
  6. [6] Tianxiang Gao, Jinzhe Li, Zhiyuan Li, Yi Chang, and Yuan Wu. More choices, fewer decisions: Ordinal-scale bias in JEV-like direct-decision models.arXiv preprint arXiv:2609.38827, 2026 on PithPith's reviewarXiv
  7. [7] Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. Recommendation as language processing (RLP): A unified pretrain, personalized prompt & predict paradigm (P5). InProceedings of the 16th ACM Conference on Recommender Systems (RecSys), 2022 on Pith
  8. [8] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. InProceedings of the 34th International Conference on Machine Learning (ICML), 2017 on Pith
  9. [9] Haoyu Han, Li Ma, Hanbing Wang, Bingheng Li, Daochen Zha, Chun How Tan, Huiji Gao, Xin Liu, Stephanie Moyerman, Sanjeev Katariya, Hui Liu, and Jiliang Tang. An embarrassingly simple graph heuristic reveals shortcut-solvable benchmarks for sequential recommendation.arXiv prepri... on PithPith's reviewarXiv
  10. [10] F. Maxwell Harper and Joseph A. Konstan. The MovieLens datasets: History and context. ACM Transactions on Interactive Intelligent Systems, 5(4), 2015 on Pith
  11. [11] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. Session-based recommendations with recurrent neural networks. InInternational Conference on Learning Representations (ICLR), 2016 on Pith
  12. [12] Sture Holm. A simple sequentially rejective multiple test procedure.Scandinavian Journal of Statistics, 6(2):65–70, 1979 on Pith
  13. [13] Yupeng Hou, Jiacheng Li, Xiangjun Fu, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. Bridging language and items for retrieval and recommendation: Benchmarking LLMs as semantic encoders.arXiv preprint arXiv:2403.03952, 2024 on PithPith's reviewarXiv
  14. [14] Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. Large language models are zero-shot rankers for recommender systems. In European Conference on Information Retrieval (ECIR), 2024. 30 on Pith
  15. [15] David R. Hunter. MM algorithms for generalized Bradley-Terry models.The Annals of Statistics, 32(1), 2004 on Pith
  16. [16] Yitong Ji, Aixin Sun, Jie Zhang, and Chenliang Li. A critical study on data leakage in recommender system offline evaluation.arXiv preprint arXiv:2010.11060, 2020 on PithPith's reviewarXiv
  17. [17] Yangqin Jiang, Xubin Ren, Lianghao Xia, Da Luo, Kangyi Lin, and Chao Huang. Recgpt: A foundation model for sequential recommendation. InProceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025. arXiv:2506.06270 on PithPith's reviewarXiv
  18. [18] Wang-Cheng Kang and Julian McAuley. Self-attentive sequential recommendation. InIEEE International Conference on Data Mining (ICDM), 2018 on Pith
  19. [19] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. LightGBM: A highly efficient gradient boosting decision tree. InAdvances in Neural Information Processing Systems 30, pages 3146–3154, 2017 on Pith
  20. [20] Kev contributors. Kev: open decision models on Qwen backbones.https://github.com/ jaredpalmer/kev, 2026. Accessed 2026-10-01 on Pith
show 24 more
  1. [21] Walid Krichene and Steffen Rendle. On sampled metrics for item recommendation. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2020 on Pith
  2. [22] Dong-Ho Lee, Adam Kraft, Long Jin, Nikhil Mehta, Taibai Xu, Lichan Hong, Ed H. Chi, and Xinyang Yi. Star: A simple training-free approach for recommendations using large language models.arXiv preprint arXiv:2410.16458, 2024 on PithPith's reviewarXiv
  3. [23] Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. LLaRA: Large language-recommendation assistant. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2024 on Pith
  4. [24] R. Duncan Luce.Individual Choice Behavior: A Theoretical Analysis. Wiley, 1959 on Pith
  5. [25] Malte Ludewig and Dietmar Jannach. Evaluation of session-based recommendation algo- rithms.User Modeling and User-Adapted Interaction, 2018 on Pith
  6. [26] Hanjia Lyu and Yinglong Xia. Decision-oriented recommendation reranking: An empirical study of its promise and limitations.arXiv preprint arXiv:2609.40241 (v2, 12 October 2026), 2026 on PithPith's reviewarXiv
  7. [27] Daniel McFadden. Conditional logit analysis of qualitative choice behavior. In Paul Zarembka, editor,Frontiers in Econometrics. Academic Press, 1974 on Pith
  8. [28] Yohei Nakajima. The log is the agent: Event-sourced reactive graphs for auditable, forkable agentic systems.arXiv preprint arXiv:2605.21997, 2026 on PithPith's reviewarXiv
  9. [29] R. L. Plackett. The analysis of permutations.Journal of the Royal Statistical Society. Series C (Applied Statistics), 24(2):193–202, 1975 on Pith
  10. [30] Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheswaran Sathiamoorthy. Recommender systems with generative retrieval. InAdvances in Neural Information Proc... on Pith
  11. [31] Mauricio Sadinle, Jing Lei, and Larry Wasserman. Least ambiguous set-valued classifiers with bounded error levels.Journal of the American Statistical Association, 114(525), 2019. 31 on Pith
  12. [32] Mark D. Smucker, James Allan, and Ben Carterette. A comparison of statistical significance tests for information retrieval evaluation. InProceedings of the 16th ACM Conference on Information and Knowledge Management (CIKM), pages 623–632, 2007 on Pith
  13. [33] Giuseppe Spillo, Alessandro Petruzzelli, Cataldo Musto, Marco de Gemmis, Pasquale Lops, and Giovanni Semeraro. Binge watch: Reproducible multimodal benchmarks datasets for large-scale movie recommendation on MovieLens-10M and 20M, 2026. Dataset, CC BY 4.0 on Pith
  14. [34] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. InProceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM), 2019 on Pith
  15. [35] TypeSafeAI. Jev: Systemonemodels.productdocumentation. https://docs.typesafe.ai, on Pith
  16. [36] Hongwei Wang, Fuzheng Zhang, Jialin Wang, Miao Zhao, Wenjie Li, Xing Xie, and Minyi Guo. RippleNet: Propagating user preferences on the knowledge graph for recommender systems. InProceedings of the 27th ACM International Conference on Information and Knowledge Management (CIKM), 2018 on Pith
  17. [37] Junting Wang, Praneet Rathi, and Hari Sundaram. A pre-trained zero-shot sequential recommendation framework via popularity dynamics. InProceedings of the 18th ACM Conference on Recommender Systems, 2024. doi: 10.1145/3640457.3688145 on Pithdoi:10.1145/3640457.3688145
  18. [38] Xiang Wang, Xiangnan He, Yixin Cao, Meng Liu, and Tat-Seng Chua. KGAT: Knowledge graph attention network for recommendation. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2019 on Pith
  19. [39] Zhaohui Wang. The recall ceiling of llm recommendation reranking. InProceedings of the ACM International Conference on Information and Knowledge Management (CIKM), 2026. arXiv:2609.27953 on PithPith's reviewarXiv
  20. [40] Zhefan Wang, Yuanqing Yu, Wendi Zheng, Weizhi Ma, and Min Zhang. MACRec: a multi- agent collaboration framework for recommendation. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2024 on Pith
  21. [41] Wikimedia Foundation. Wikidata. https://www.wikidata.org, 2026. Data under CC0; queried through the QLever mirror, October 2026 on Pith
  22. [42] Zhenrui Yue, Sara Rabhi, Gabriel de Souza Pereira Moreira, Dong Wang, and Even Oldridge. Llamarec: Two-stage recommendation using large language models for ranking.arXiv preprint arXiv:2311.02089, 2023 on PithPith's reviewarXiv
  23. [43] Jiamu Zhang, Tianze Yang, Yucheng Shi, Evan Chen, Zixiang Nie, Kelly Wan, Liangjie Hong, Ninghao Liu, and Liang Wu. AnyJev technical report.arXiv preprint arXiv:2610.00831, 2026 on PithPith's reviewarXiv
  24. [44] Kaiji Zhou, Vladimir Kalmykov, and Yue Feng. RouteRec: Strict evaluation of recommender- agent selection and aggregation. InAgentSearch 2026, the First Workshop on Indexing, Retrieval, and Ranking of AI Agents, co-located with SIGIR 2026, 2026. arXiv:2607.09908. 32 A Additiona... 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.

No comments yet.

Sign in to comment