Pith.

Information Retrieval · arXiv:2611.02487 v1 · submitted 2 Nov 2026 · 43 references

RecRSI: Mechanism-Guided Discovery of Recommendation Algorithms

Tingran Gan, Zhouziyue Lu, Dian Jin, Tao Li, Yang Gao, Ming Liu, Weiwen Liu, Jianghao Lin

show 2 more authors

Weinan Zhang, Yong Yu

The paper says

“Improving recommendation algorithms for the Web requires coordinating interacting design choices and learning from experiments. We present RecRSI, an LLM-based research framework that uses mechanism hypotheses to organize algorithm discovery.”

more of the abstract

Each hypothesis links a proposed mechanism change to a reference design, a rationale, and an expected observation. MechSpace derives implementable primitives and composition rules from the task source; Research Line uses this space and prior evidence to develop and test hypotheses; Mechanism Feedback carries outcomes and their design context into subsequent research. This organization makes each experiment inform which mechanisms to develop, revise, or combine next. Across four initial recommenders and four datasets, RecRSI yields average relative improvements of 8.48% in NDCG@5 and 6.88% in HR@5 over the initial algorithms, and 2.60% and 2.20%, respectively, over the strongest competing result for each setting and metric. Component and process studies link these gains to coordinated mechanism choices and evidence-guided follow-up. Our code is publicly available at https://github.com/DynamisG210/RecRSI.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 3 Nov 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, with conditions. A serious LLM-discovery framework whose mechanism-hypothesis organization is a real idea and whose evaluation is unusually careful - but the headline 16-setting averages mix 8 searched and 8 transferred settings, so the abstract overstates the uniformity of the comparison. This line is wrong

3 major objections · 4 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 43 references say.

The paper says, at Abstract, §4.1, Appendices A–B

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The 16-setting headline averages conflate two different claims. Appendix A states that Table 1 combines 'task-specific search winners and source-selected designs retrained on target data,' and Appendix B defines the common search cohort as only eight tasks (SASRec/CDs; MoRec/CDs, Office, MovieLens-1M; BERT4Rec and NGCF each on Electronics and MovieLens-1M).

more of objection 1

The other eight Table 1 settings therefore measure transfer quality of source-search winners, not discovery conducted under the stated per-setting 12-evaluation budget. The abstract and §4.1 present the 8.48%/6.88% and 2.60%/2.20% averages over all 16 settings as the central result, and the paper nowhere reports the margins restricted to the eight directly searched settings, nor a confidence interval for that subset. The bootstrap in Appendix A is described as being over 'the eight MoRec and BERT4Rec settings,' which is not obviously the same set as the Appendix B cohort; the paper currently contains at least three different 'eight' sets (Figure 3 cohort, Appendix A bootstrap, Appendix B cohort) without saying how they relate. The disclosure in Appendix A shows this is not a fabrication concern, but it does not resolve the load-bearing question: whether the headline gains survive when the 'strongest competing result' is required to come from a search on that same setting. Please label each Table 1 row as direct-search or transfer, report the headline margins on the direct-search subset with bootstrap intervals, and state in §4.1 that the other eight settings assess design reuse rather than discovery.

The paper says, at Appendix B, Table 10

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The claim that all six systems were compared under a shared 12-evaluation budget is only partially supported. Appendix B reports that a pre-specified USD 7–8 API safeguard stopped 18 DeepEvolve or Self-EvolveRec searches, and Table 10 shows several repeats with very few completed evaluations (e.g., DeepEvolve MoRec/CDs R3 = 0, MoRec/Office R1 = 0 and R2 = 1, Self-EvolveRec NGCF/Electronics R2 = 0).

more of objection 2

Because Table 1 entries are means over three repeats, margins over those systems partly compare RecRSI against searches that consumed far less than the nominal evaluation budget. The paper's mitigation is real but incomplete: the safeguard is described as fixed and symmetric, and Figure 3(c) provides a common-token-ceiling check on eight tasks, which is the right kind of control. What is missing is a per-system, per-task distribution of completed evaluations on the eight-task cohort, a statement of how the USD 7–8 safeguard was chosen relative to the expected cost of 12 completed evaluations, and a verification that the 'strongest competing result' margins are unchanged when the comparison is restricted to repeats that actually completed the full evaluation budget. Without that, a reader cannot tell how much of the average margin is a budget artifact.

The paper says, at §4.1, Table 1

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The sentence 'RecRSI improves on the initial algorithm in all 16 settings and achieves the highest mean NDCG@5 in 11 settings and the highest mean HR@5 in 12' is not reproducible from the printed table. From the rounded means in Table 1, RecRSI is the best system in 12 NDCG@5 settings (losing to the strongest printed competitor on SASRec/MovieLens-1M, MoRec/Office, MoRec/MovieLens-1M, and BERT4Rec/MovieLens-1M) and in 13 HR@5 settings (losing on SASRec/Office, MoRec/Office, and BERT4Rec/MovieLens-1M).

more of objection 3

The text's counts presumably come from unrounded search-level means, but those are not reported anywhere; the discrepancy should be resolved by publishing the unrounded per-repeat values or by correcting the counts. This matters because the '11/12' sentence is the paper's per-setting summary of the same evidence that produces the headline averages.

4 minor objections

The paper says, at §4.1, Table 7

Pith says: minor objection 1 of 4

Several advertised per-setting improvements are smaller than the search-to-search standard deviations in Table 7; for example, the SASRec/Electronics NDCG@5 gain over seed is 0.0107 while RecRSI's between-search SD is 0.0185, and the MoRec/Office HR@5 gain is 0.0009 while the relevant SDs are about 0.004–0.005. The grouped bootstrap is the appropriate level of inference; please add a sentence telling readers that per-setting winner statements are based on mean values and should not be read as individually significant.

The paper says, at Appendix B, Figure 3

Pith says: minor objection 2 of 4

The caption of Figure 3 says '(a,b) Eight-task final test NDCG@5 gain over Seed versus API reference cost (USD) and per GPU-hour,' which is ambiguous about which panel corresponds to which quantity; please split the caption or name the panels explicitly. Also, the claim that the eight tasks in Figure 1(b) and Figure 3 'are the same eight tasks' should be cross-referenced to the explicit task list in Appendix B at first use.

show 2 more

The paper says, at Appendix D

Pith says: minor objection 3 of 4

For reproducibility, the LLM configuration should be more precise: 'temperature left unset' should state the actual default temperature used by the API, and the exact model identifiers (e.g., gpt-5.6-terra versus gpt-5.6-luna, and the reasoning configuration 'medium') should be given with access dates, since all results depend on these settings.

The paper says, at §4.3.3, Table 5

Pith says: minor objection 4 of 4

The comparison between the automatically generated MechSpace (A) and the human-designed space (H) on SASRec/CDs is interesting, but the construction of H is described only as the authors 'manually specifying mechanisms, parameter domains, and composition rules'; since the H space is a home-grown baseline, please clarify who designed it and whether it was intended as a strong or minimal human baseline, so readers can calibrate the 19.65% difference.

What would settle it

Recompute the headline margins on the eight Appendix B direct-search settings only: SASRec/CDs; MoRec/CDs, Office, MovieLens-1M; BERT4Rec/Electronics, MovieLens-1M; NGCF/Electronics, MovieLens-1M. Label each of the 16 Table 1 settings as direct search or source-transfer from the search lineage, compute mean relative NDCG@5 and HR@5 gains over the strongest baseline mean per setting, and bootstrap over the three search repeats for the eight direct settings. If the direct-subset 95% confidence interval excludes the reported 2.60% / 2.20% margins or falls materially below them, the abstract should state the direct-search-subset result separately from the transfer-inclusive 16-setting average.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 3 Nov 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 30 s · 21,906 tokens read, 13,412 written
  3. Editor wrote the verdict above 1 min 2 s · 22,938 tokens read, 12,459 written
  4. Referee major revision: 3 major objections, 4 minor 2 min 15 s · 22,324 tokens read, 22,404 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

Three things worth knowing up front. The organizing idea is real: a mechanism hypothesis - a proposed change tied to a reference design, a rationale, and an expected observation - as the unit that carries LLM-guided discovery forward instead of raw program edits. The empirical work is among the most careful in this line: 16 settings, three search repeats, shared splits and protocols, bootstrap intervals, ablations, and a human-vs-generated design space comparison. And the headline numbers are softer than the abstract makes them look, because they mix searched and transferred settings.

What's good. The paper tests its own thesis in ways that mostly land. Table 11 is the standout: layer aggregation hurts without label smoothing (-0.0014) and helps with it (+0.0015), which is the "mechanisms interact" claim turned into a measurement. The space-feedback interaction (+0.6064 pp) in Table 2 is a clean ablation result. The history experiments (with vs. reset: 0.3947 vs. 0.3681) support the cumulative-evidence story. Table 5's generated MechSpace beating a human-designed space by ~19.6% is a nice control. The disclosures are honest, if tucked into appendices: the searched/transferred mix, 18 API-safeguard-stopped baseline searches, OpenEvolve's seed-equivalent zero-SD rows.

The concern that matters. The stress-test note is right. Appendix A does disclose that the 16 settings combine "task-specific search winners and source-selected designs retrained on target data," and Appendix B defines the shared 12-evaluation cohort as eight tasks. So the abstract's averages blend 8 searches and 8 transfers, and the pure direct-search margin over the strongest competitor is never reported with a confidence interval. The bootstrap covers a different, partially mixed subset (the eight MoRec and BERT4Rec settings). This does not refute the claim - Figure 3's eight-task curves still favor RecRSI (7.36% vs. 4.68% test NDCG@5 gain over seed vs. RecHarness), and the leave-one-setting-out analysis stays positive - but the paper should put the eight-task restricted margins and their CI in the main text. As written, the headline overstates how uniform the comparison is.

Secondary. Baseline budget parity is real but not central: DeepEvolve and Self-EvolveRec repeats stopped at 0-6 evaluations by the API safeguard; disclosed and symmetric, and the strongest competitor (RecHarness) was unaffected, so the main margin stands. Exact numbers depend on a proprietary LLM and an unverified repo, so independent reproduction is impossible - a property of the field more than a defect here. Minor: no limitations section in the main text; the candor is all in appendices.

Citations are fine; related work is credited, including same-group surveys. No fabrication red flags, no circularity. This paper deserves a serious referee, and the revision is tractable: report the direct-search decomposition, the restricted CI, and a budget-parity caveat. I'd bring it to a reading group and would cite the framework.

The strongest objection

The central claim that mechanism-guided search outperforms program-level baselines under a shared experimental budget rests on the 16-setting averages of 8.48% / 6.88% over seeds and 2.60% / 2.20% over the strongest competitor. But the paper's own scoping statements show that Table 1 is not a 16-setting search comparison. Appendix A says it evaluates 16 settings by 'combining task-specific search winners and source-selected designs retrained on target data,' while Appendix B defines the shared 12-evaluation cohort as only eight tasks: SASRec/CDs; MoRec/CDs, Office, MovieLens-1M; and BERT4Rec and NGCF each on Electronics and MovieLens-1M. The other eight Table 1 settings therefore measure transfer of source-search winners, not discovery conducted under the stated per-setting budget. The paper never reports the headline margins restricted to the eight directly searched settings, nor a confidence interval for those eight, so the abstract's average conflates two different claims: quality of mechanism-guided search and quality of transferred designs. This is not a fabrication concern, and the Appendix A sentence shows the authors disclose the mix, but the disclosure does not resolve the load-bearing question: whether the 8.48% / 2.60% numbers survive when 'strongest competing result' is required to be a result produced by a search on that same setting. My rough reading of Table 1 suggests the direct-search subset may still be positive, so the issue is one of precision and scope rather than a demonstrated falsehood, but it is the condition that actually secures the central claim. The 18 API-safeguard-stopped baseline searches are documented and mostly did not define the strongest competitor, so I do not treat them as the primary attack; the missing direct-search decomposition is more structural.

The strongest claim is that mechanism-guided search produces better discovered recommenders than program-level evolution under a shared experimental budget. The reader's weakest assumption was baseline parity, focusing on API-safeguard-stopped searches and OpenEvolve's seed-equivalent results. Those are real concerns, but the paper's own Table 10 shows the stopped runs are often not the strongest competitor in their settings, so that concern is less decisive than the structural issue of what the 16-setting aggregate actually measures. The more load-bearing condition is that the headline averages are computed over settings with two different evidentiary statuses: eight where each system searched under the shared 12-evaluation ceiling, and eight where scores come from source-selected designs retrained on target data. The paper discloses this in Appendix A, but never reports the direct-search-only decomposition, so the central percentage cannot be audited from the tables alone. The existing bootstrap analysis covers eight MoRec and BERT4Rec settings, which includes some transfer settings and excludes SASRec and NGCF, so it does not settle the direct-search subset question. This concern reinforces the CONDITIONAL verdict rather than overturning it: the empirical claim is plausible and broadly reported, but the headline should be verified against the directly searched subset before it is read as evidence about search under a fixed budget. I credit the paper for documenting safeguard stops, reporting search-level variability on a subset, and providing a public code link, but none of these replaces the missing direct-search decomposition.

The referee's report

If the central empirical claim holds, this is a meaningful systems contribution: it is one of the first demonstrations in recommender-systems research that organizing LLM-driven search around mechanism-level hypotheses outperforms program-level evolution baselines under shared evaluation and cost budgets. The empirical design is stronger than typical for this literature and deserves explicit credit: publicly released code; shared data splits and evaluation protocols across all systems; three independent search repeats; held-out final test evaluation with dev-set selection; bootstrap intervals on a grouped subset; leave-one-out robustness checks over settings, models, and datasets; component comparisons with fixed seeds; and a common-token-budget comparison in addition to the nominal 12-evaluation ceiling. The process studies (continuation vs. restarts, history-reset comparisons, and frozen-decision feedback comparisons) give the paper's central mechanism claim more internal support than a purely aggregate comparison would. The main weaknesses are in how the headline numbers are scoped and reported, not in the execution of the experiments themselves.

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

RecRSI claims that automated algorithm discovery works better when the search unit is a mechanism hypothesis rather than a raw program edit. Each hypothesis packages a proposed design change together with the reference design it builds on, the rationale for trying it, and the observation it expects, so that every experiment, success or failure, becomes evidence about which mechanisms to keep, revise, or combine next. Over 16 settings spanning four starting recommenders and four datasets, the framework reports average relative gains of 8.48% in NDCG@5 and 6.88% in HR@5 over the initial algorithms, and 2.60% and 2.20% over the strongest competing system in each setting and metric. If this holds, it matters because research budgets, not model capability, are what limit automated discovery: organizing experiments as explicit testable questions yields better algorithms than more evolution steps at the same evaluation cost.

What carries the argument. The load-bearing object is the mechanism hypothesis, a four-part record $\mathcal{h}_t = (C^{\mathrm{ref}}_t, C_t, \rho_t, \epsilon_t)$ that ties a proposed mechanism change to the reference composition it modifies, the rationale for the intervention, and the observation it is expected to produce. Around it sit MechSpace, the task-derived registry of primitives and composition rules, and Mechanism Feedback, which stores each outcome with its design context so later hypotheses can build on successful designs and use failed ones to change direction. The paper's worked example shows the kind of interaction this exposes: a cosine-temperature scoring change rewrites `Model.predict_score`, the function shared by training loss and candidate evaluation, so a paired-margin objective added through the same scores makes the compatibility of training and ranking an explicit, testable relationship rather than a hidden coupling.

Core claim. The central claim is that making the mechanism hypothesis the organizing unit of research converts a search loop into a cumulative experimental program. A hypothesis $\mathcal{h}_t = (C^{\mathrm{ref}}_t, C_t, \rho_t, \epsilon_t)$ names the reference composition, the proposed composition, the rationale, and the expected observation; MechSpace turns the task source into a validated space of primitives and composition rules so the hypothesis is implementable; Research Line fills a proposal pool through four roles (composition, lineage refinement, falsification design, frontier exploration); and Mechanism Feedback writes each outcome back into the shared research state together with the changed and unchanged mechanisms. The paper reports that across 16 model–dataset settings RecRSI improves on every initial recommender, reaches the highest mean NDCG@5 in 11 settings and the highest mean HR@5 in 12, and averages 8.48% and 6.88% relative gains over the seeds and 2.60% and 2.20% over the strongest competitor per setting. Component studies trace the gains to coordinated mechanism combinations, with scoring and objective changes co-occurring in seven of eight BERT4Rec designs, and to hypothesis-linked feedback, which outperforms generic experiment summaries in seven of ten held decision points.

The premise it rests on. The central claim stands on the premise that all six systems ran on truly equal footing — same data splits and protocol, same 12-evaluation ceiling, and cost safeguards that did not systematically truncate baseline searches before they could improve; if the baselines were cost-starved, the reported advantage measures budget accounting rather than mechanism organization.

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

At a fixed ceiling of 12 genuine evaluations, mechanism-guided search reaches higher discovery quality than the strongest program-evolution baseline, implying that research structure, not raw search breadth, drives the gains.

Continued searches find new best candidates after evaluation 12 in every continuation block, so the framework's value grows with budget rather than saturating at the short-search horizon.

Access to accumulated research history improves successive candidate quality (mean NDCG@5 of 0.3947 versus 0.3681), so the mechanism evidence itself, not just the current best program, is a reusable research asset.

The most effective discovered designs combine scoring and objective changes that share the same score function, suggesting mechanism-guided search is particularly valuable where training and prediction interact.

At a common 3-million-token budget, RecRSI reaches a mean validation gain of 7.72% versus 5.09% for the strongest competitor, so the advantage is not bought simply by spending more LLM tokens.

Editorial extension: the mechanism-hypothesis template of change, reference design, rationale, and expected observation transfers naturally to automated experimentation outside recommendation, wherever an experiment's interpretation depends on what was held fixed.

Editorial extension: the per-setting edge over the strongest competitor ranges from roughly +12% to about −3%, with the negative cases clustered on MovieLens-1M and NGCF/Electronics, so a useful follow-up would identify the task properties that predict when mechanism-guided search pays off.

Editorial extension: because the API safeguard halted 18 DeepEvolve or Self-EvolveRec searches before their evaluation ceiling, an independent replication with matched cost headroom is the cleanest test of whether the 2.60%/2.20% advantage is a property of the method or of budget accounting.

Editorial extension: a variant using a smaller, cheaper LLM behind the same mechanism scaffolding would test whether the framework's value lies in its structure or in the expensive reasoning tokens spent on hypothesis formulation.

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

Improving recommendation algorithms for the Web requires coordinating interacting design choices and learning from experiments. We present RecRSI, an LLM-based research framework that uses mechanism hypotheses to organize algorithm discovery. Each hypothesis links a proposed mechanism change to a reference design, a rationale, and an expected observation. MechSpace derives implementable primitives and composition rules from the task source; Research Line uses this space and prior evidence to develop and test hypotheses; Mechanism Feedback carries outcomes and their design context into subsequent research. This organization makes each experiment inform which mechanisms to develop, revise, or combine next. Across four initial recommenders and four datasets, RecRSI yields average relative improvements of 8.48% in NDCG@5 and 6.88% in HR@5 over the initial algorithms, and 2.60% and 2.20%, respectively, over the strongest competing result for each setting and metric. Component and process studies link these gains to coordinated mechanism choices and evidence-guided follow-up. Our code is publicly available at https://github.com/DynamisG210/RecRSI.

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 · 2026-11-03

The authors' main claim, as the classifier picked it out 0.92

“Across four initial recommenders and four datasets, RecRSI yields average relative improvements of 8.48% in NDCG@5 and 6.88% in HR@5 over the initial algorithms, and 2.60% and 2.20%, respectively, over the strongest competing result for each setting and metric.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.08 · human participants 0.08 · controlled comparison 0.90 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.97 · 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
Improving recommendation algorithms for the Web requires coordinating interacting design choices and learning from experiments. background 1.00 0.06 0.03 asserted 0.99
We present RecRSI, an LLM-based research framework that uses mechanism hypotheses to organize algorithm discovery. contribution 0.97 0.90 0.07 asserted 1.00
Each hypothesis links a proposed mechanism change to a reference design, a rationale, and an expected observation. method 0.94 0.69 0.03 asserted 1.00
MechSpace derives implementable primitives and composition rules from the task source; Research Line uses this space and prior evidence to develop and test hypotheses; Mechanism Feedback carries outcomes and their design context into subsequent research. method 0.99 0.84 0.04 asserted 0.99
This organization makes each experiment inform which mechanisms to develop, revise, or combine next. method 0.46 0.57 0.14 asserted 1.00
Across four initial recommenders and four datasets, RecRSI yields average relative improvements of 8.48% in NDCG@5 and 6.88% in HR@5 over the initial algorithms, and 2.60% and 2.20%, respectively, over the strongest competing result for each setting and metric. result 1.00 0.97 0.55 asserted 1.00
Component and process studies link these gains to coordinated mechanism choices and evidence-guided follow-up. interpretation 0.97 0.89 0.23 asserted 1.00
Our code is publicly available at https://github.com/DynamisG210/RecRSI. contribution 0.90 0.66 0.02 asserted 0.98

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98

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, 2026-11-03: 1 request, 6,227 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 (4)
Figure 1 of the paper
Figure 1. Research organization and algorithm discovery. (a) Candidate-level evolution organizes iteration around programs;
Figure 2 of the paper
Figure 2. Overview of RecRSI. MechSpace supplies the design space for mechanism hypotheses; Research Line develops them
Figure 3 of the paper
Figure 3. Discovery quality and search cost. (a,b) Eight-task
Figure 4 of the paper
Figure 4. Overlaid research progress on MoRec/CDs. Colors

Figures are the authors', reproduced from the paper.

References (43)
  1. [1] Yongqiang Chen, Chenxi Liu, Zhenhao Chen, Tongliang Liu, Bo Han, and Kun Zhang. 2026. CausalEvolve: Towards Open-Ended Discovery with Causal Scratch- pad. arXiv:2603.14575 doi:10.48550/arXiv.2603.14575 on Pithdoi:10.48550/arxiv.2603.14575
  2. [2] Maurizio Ferrari Dacrema, Paolo Cremonesi, and Dietmar Jannach. 2019. Are We Really Making Much Progress? A Worrying Analysis of Recent Neural Rec- ommendation Approaches. InProceedings of RecSys. ACM, New York, NY, USA, 101–109. doi:10.1145/3298689.3347058 on Pithdoi:10.1145/3298689.3347058
  3. [3] F. Maxwell Harper and Joseph A. Konstan. 2015. The MovieLens Datasets: History and Context.ACM Transactions on Interactive Intelligent Systems5, 4, Article 19 (2015), 19 pages. doi:10.1145/2827872 on Pithdoi:10.1145/2827872
  4. [4] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. InProceedings of SIGIR. ACM, New York, NY, USA, 639–648. doi:10.1145/3397271.3401063 on Pithdoi:10.1145/3397271.3401063
  5. [5] Seokhoon Jeong, Mijung Kim, and Taehwan Kim. 2026. Agentic Neural Architec- ture Search. arXiv:2607.07984 doi:10.48550/arXiv.2607.07984 on PithPith's reviewarXivdoi:10.48550/arxiv.2607.07984
  6. [6] Zhengyao Jiang, Dominik Schmidt, Dhruv Srikanth, Dixing Xu, Ian Kaplan, Deniss Jacenko, and Yuxiang Wu. 2025. AIDE: AI-Driven Exploration in the Space of Code. arXiv:2502.13138 doi:10.48550/arXiv.2502.13138 on PithPith's reviewarXivdoi:10.48550/arxiv.2502.13138
  7. [7] Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recom- mendation. In2018 IEEE International Conference on Data Mining (ICDM). IEEE, Piscataway, NJ, USA, 197–206. doi:10.1109/ICDM.2018.00035 on Pithdoi:10.1109/icdm.2018.00035
  8. [8] Andrej Karpathy. 2026.autoresearch: AI Agents Running Research on Single- GPU Nanochat Training Automatically. GitHub. https://github.com/karpathy/ autoresearch on Pith
  9. [9] Sein Kim, Sangwu Park, Hongseok Kang, Wonjoong Kim, Jimin Seo, Yeonjun In, Kanghoon Yoon, Hyunsik Jeon, and Chanyoung Park. 2026. Self-EvolveRec: Self-Evolving Recommender Systems with LLM-Based Directional Feedback. arXiv:2602.12612 doi:10.48550/arXiv.2602.12612 on PithPith's reviewarXivdoi:10.48550/arxiv.2602.12612
  10. [10] Robert Tjarko Lange, Yuki Imajuku, and Edoardo Cetin. 2026. ShinkaE- volve: Towards Open-Ended and Sample-Efficient Program Evolution. InThe Fourteenth International Conference on Learning Representations. 74026–74078. https://proceedings.iclr.cc/paper_files/paper/2026/hash/ 7... on Pith
  11. [11] Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, and Weinan Zhang. 2025. How Can Recommender Systems Benefit from Large Language Models: A Survey.ACM Transactions on Informati... on Pithdoi:10.1145/3678004
  12. [12] Haoran Ling, Yuecheng Li, Zeyu Song, Jing Yao, Shuwen Kang, Chi Lu, Wenjin Wu, and Peng Jiang. 2026. RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems. arXiv:2607.29241 https://arxiv.org/abs/ 2607.29241 on PithPith's reviewarXiv
  13. [13] Bin Liu, Chenxu Zhu, Guilin Li, Weinan Zhang, Jincai Lai, Ruiming Tang, Xi- uqiang He, Zhenguo Li, and Yong Yu. 2020. AutoFIS: Automatic Feature Inter- action Selection in Factorization Models for Click-Through Rate Prediction. In Proceedings of KDD. ACM, New York, NY, USA, 26... on Pithdoi:10.1145/3394486
  14. [14] Fei Liu, Xialiang Tong, Mingxuan Yuan, Xi Lin, Fu Luo, Zhenkun Wang, Zhichao Lu, and Qingfu Zhang. 2024. Evolution of Heuristics: Towards Efficient Auto- matic Algorithm Design Using Large Language Model. InProceedings of the 41st International Conference on Machine Learning (... on Pith
  15. [15] Gang Liu, Yihan Zhu, Jie Chen, and Meng Jiang. 2025. Scientific Algorithm Discovery by Augmenting AlphaEvolve with Deep Research. arXiv:2510.06056 doi:10.48550/arXiv.2510.06056 on Pithdoi:10.48550/arxiv.2510.06056
  16. [16] Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. 2024. The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. arXiv:2408.06292 doi:10.48550/arXiv.2408.06292 on PithPith's reviewarXivdoi:10.48550/arxiv.2408.06292
  17. [17] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. Self-Refine: It... on Pithdoi:10.52202/075280-2019
  18. [18] Kelong Mao, Jieming Zhu, Jinpeng Wang, Quanyu Dai, Zhenhua Dong, Xi Xiao, and Xiuqiang He. 2021. SimpleX: A Simple and Strong Baseline for Collaborative Filtering. InProceedings of the 30th ACM International Conference on Information and Knowledge Management (CIKM ’21). Associ... on Pithdoi:10.1145/3459637.3482297
  19. [19] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel on Pith
  20. [20] Lingyu Mu, Hao Deng, Haibo Xing, Jinxin Hu, Yu Zhang, and Xiaoyi Zeng on Pith
show 23 more
  1. [21] Alexander Novikov, Ngân V ˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Push... on PithPith's reviewarXivdoi:10.48550/arxiv.2506.13131
  2. [22] Esteban Real, Chen Liang, David So, and Quoc Le. 2020. AutoML-Zero: Evolving Machine Learning Algorithms From Scratch. InProceedings of ICML (Proceedings of Machine Learning Research, Vol. 119). PMLR, Cambridge, MA, USA, 8007–8019. https://proceedings.mlr.press/v119/real20a.html on Pith
  3. [23] Steffen Rendle, Walid Krichene, Li Zhang, and John Roberts Anderson. 2020. Neural Collaborative Filtering vs. Matrix Factorization Revisited. InProceedings of the 14th ACM Conference on Recommender Systems. Association for Computing Machinery, New York, NY, USA, 240–248. doi:1... on Pithdoi:10.1145/3383313.3412488
  4. [24] Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi on Pith
  5. [25] Yuiko Sakuma, Masakazu Yoshimura, Marcel Gröpl, Zitang Sun, Junji Otsuka, Atsushi Irie, and Takeshi Ohashi. 2026. Structuring Open-Ended NAS: Semi- Automated Design Knowledge Structuring with LLMs for Efficient Neural Archi- tecture Search. arXiv:2605.19247 doi:10.48550/arXiv.... on PithPith's reviewarXivdoi:10.48550/arxiv.2605.19247
  6. [26] Asankhaya Sharma. 2025.OpenEvolve: An Open-Source Evolutionary Coding Agent. GitHub. https://github.com/algorithmicsuperintelligence/openevolve on Pith
  7. [27] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language Agents with Verbal Re- inforcement Learning. InAdvances in Neural Information Processing Systems, Vol. 36. Curran Associates, Inc., Red Hook, NY, USA, 8634– on Pith
  8. [28] Qingquan Song, Dehua Cheng, Hanning Zhou, Jiyan Yang, Yuandong Tian, and Xia Hu. 2020. Towards Automated Neural Interaction Discovery for Click- Through Rate Prediction. InProceedings of KDD. ACM, New York, NY, USA, 945–955. doi:10.1145/3394486.3403137 on Pithdoi:10.1145/3394486.3403137
  9. [29] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang on Pith
  10. [30] Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural Graph Collaborative Filtering. InProceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval. Association for Computing Machinery, New York, NY, U... on Pith
  11. [31] Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, and David Ha. 2025. The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search. arXiv:2504.08066 doi:10. 48550/arXiv.2504.08066 on PithPith's reviewarXivdoi:10.48550/arxiv.2504.08066
  12. [32] Shuang Yang, Zijie Zhuang, Changxin Lao, Pengbo Xu, Hanwen Xu, Yusheng Huang, Han Gao, Guanchen Wang, Tianbao Ma, Linxun Chen, Peilin Song, Xuming Wang, Chen Li, Fan Wu, Tao Wang, Zibo Zhao, Xiangyu Wu, An Liu, Fei Pan, Peng Jiang, Chen Yang, Zhaojie Liu, and Wenwu Ou. 2026. A... on PithPith's reviewarXiv
  13. [33] Haoran Ye, Jiarui Wang, Zhiguang Cao, Federico Berto, Chuanbo Hua, Haeyeon Kim, Jinkyoo Park, and Guojie Song. 2024. ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution. InAdvances in Neural Infor- mation Processing Systems, Vol. 37. Curran Associates, I... on Pith
  14. [34] Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. 2023. Where to Go Next for Recommender Systems? ID- vs. Modality-based Recommender Models Revisited. InProceedings of the 46th International ACM SIGIR Conference on Research and Devel... on Pithdoi:10.1145/3539618.3591932
  15. [35] Xingshan Zeng, Zishan Xu, Boju Zhang, Yuzhou Wu, Lingzhi Wang, Jianghao Lin, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Weinan Zhang, Yong Yu, Qun Liu, and Weiwen Liu. 2026. What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents. arXiv:2608.27260 ... on PithPith's reviewarXivdoi:10.48550/arxiv.2608.27260
  16. [36] Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune. 2025. Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents. arXiv:2505.22954 doi:10.48550/arXiv.2505.22954 on PithPith's reviewarXivdoi:10.48550/arxiv.2505.22954
  17. [37] Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, Yingqian Min, Zhichao Feng, Xinyan Fan, Xu Chen, Pengfei Wang, Wendi Ji, Yaliang Li, Xiaoling Wang, and Ji-Rong Wen. 2021. RecBole: Towards a Unified, Com... on Pith
  18. [38] Chenyu Zhou, Huacan Chai, Wenteng Chen, Zihan Guo, Rong Shan, Yuanyi Song, Tianyi Xu, Yingxuan Yang, Aofan Yu, Weiming Zhang, Congming Zheng, Jiachen Zhu, Zeyu Zheng, Zhuosheng Zhang, Xingyu Lou, Changwang Zhang, Zhihui Fu, Jun Wang, Weiwen Liu, Jianghao Lin, and Weinan Zhang.... on PithPith's reviewarXivdoi:10.48550/arxiv.2604.08224
  19. [2015] Image-based Recommendations on Styles and Substitutes. InProceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval. Association for Computing Machinery, New York, NY, USA, 43–52. doi:10.1145/2766462.2767755 on Pithdoi:10.1145/2766462.2767755
  20. [2019] BERT4Rec: Sequential Recommendation with Bidirectional Encoder Rep- resentations from Transformer. InProceedings of the 28th ACM International Conference on Information and Knowledge Management. Association for Comput- ing Machinery, New York, NY, USA, 1441–1450. doi:10.1145/3... on Pithdoi:10.1145/3357384.3357895
  21. [2024] Mathematical Discoveries from Program Search with Large Language Models.Nature625 (2024), 468–475. doi:10.1038/s41586-023-06924-6 on Pithdoi:10.1038/s41586-023-06924-6
  22. [2026] EvoRec: Self Evolving Agentic Recommender Systems. arXiv:2606.28368 doi:10.48550/arXiv.2606.28368 Gan et al on PithPith's reviewarXivdoi:10.48550/arxiv.2606.28368
  23. [8652] https://proceedings.neurips.cc/paper_files/paper/2023/hash/ 1b44b878bb782e6954cd888628510e90-Abstract-Conference.html on Pith

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