Pith. sign in

REVIEW 3 major objections 6 minor 86 references

RecHarness claims that splitting edit-direction choice from code writing — a bandit router chooses the direction, the LLM writes the edit — yields more stable, budget-efficient recommender optimization than letting the LLM do both.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 10:56 UTC pith:XAU44Q2K

load-bearing objection Solid system paper with a real, if incremental, idea: decoupling edit-direction selection (bandit) from code mutation (LLM) buys stability and budget efficiency; but the 'first' claim and missing uncertainty quantification need fixing. the 3 major comments →

arxiv 2607.29241 v1 pith:XAU44Q2K submitted 2026-07-31 cs.IR cs.AIcs.CL

RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems

Haoran Ling , Yuecheng Li , Zeyu Song , Jing Yao , Shuwen Kang , Chi Lu , Wenjin Wu , Peng Jiang This is my paper
classification cs.IR cs.AIcs.CL
keywords recommender systemsautomated model optimizationmulti-armed banditsThompson samplingLLM agentsself-evolving recommender systemscode generationonline A/B test
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper's central claim is that the instability of LLM-driven recommender optimization comes from letting a single model both decide what to change and write the change. RecHarness splits the job: a Thompson-sampling bandit router allocates a limited trial budget across human-defined edit directions using validation outcomes, and the LLM only turns the chosen direction into a concrete hypothesis and executable code edit. The authors argue — and support with offline experiments across two recommendation tasks and eight model backbones, plus a 7-day online A/B test — that this routing makes search more stable and uses scarce trials more effectively than letting the LLM freely decide where to explore. If true, automated recommender iteration becomes reliable enough for industrial deployment, as long as the space of possible changes is human-specified.

Core claim

The discovery is a control split. RecHarness frames recommender model iteration as an incumbent-based bandit problem: a fixed set of human-defined 'edit arms' (learning-rate schedules, sequence pooling, loss functions, structural jumps such as encoder upgrades) each carries a Beta posterior over binary validation success. Each round, Thompson sampling selects the arms with the highest sampled success probability; the LLM then proposes a hypothesis and code mutation within the selected arm. Success is binary and normalized within a parallel trial group. When recent improvement stagnates, a jump-basin flag admits structural 'jump arms,' but a jump is accepted only if, after a retuning window,

What carries the argument

The central object is the edit-arm set A = A_local ∪ A_jump, human-defined interpretable edit dimensions rather than exact patches or hyperparameters. The bandit router maintains a Beta(α, β) posterior per arm, updated with a binary success signal (valid, above the group average, and not below the incumbent); Thompson sampling draws from these to choose which arms receive the next parallel trial group. Experiment Skill — a distilled textual memory of successes, failures, and trends — conditions the LLM's mutation operator but never updates the posterior, so scalar evidence and textual reasoning play separate roles. The jump-basin mechanism measures the improvement rate over a window and only

Load-bearing premise

The genuinely useful modification must be expressible as one of the human-predefined edit arms; if the best improvement is not in the list, the harness can never propose it, no matter how good the routing and the LLM are.

What would settle it

Run the same ablation budget on a new backbone with an arm set that omits the structural jump arm (the one that produced the online win); if restricting the arms collapses RecHarness's advantage over the LLM-only baseline, the gains are attributable to the arm set, not the routing. Alternatively, give the TR w/ LLM variant the same Experiment Skill and the same arm names as hints; if it then matches RecHarness's ~48% trial win rate, Thompson routing is not doing the work.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, recommender optimization reduces to a budgeted bandit problem over interpretable edit dimensions, not open-ended code search.
  • LLM reasoning is most valuable inside a constrained direction; using it for exploration decisions wastes trial budget.
  • The binary, group-normalized success signal lets the router rank directions without being fooled by scale differences across parallel trials.
  • The jump-basin gate allows structural change to be evaluated after local adaptation, so architecture upgrades are not rejected on their first-step score.
  • A 7-day online test shows a structural change discovered this way can transfer to business metrics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Not drawn in the paper: the ceiling of the whole approach is the human-curated edit-arm space; the winning online edit only existed because a human had pre-defined a structural jump arm, so performance is upper-bounded by the arm designer's foresight.
  • The same separation — bandit direction selection plus LLM generation — could transfer to other expensive iterative loops, such as data pipeline tuning or kernel optimization, where each trial is costly and validity is noisy; the paper only demonstrates it for recommender training.
  • The ablation TR w/ LLM suggests that textual feedback alone is not enough; an implication the authors leave implicit is that an LLM's reading of experiment logs is not a substitute for explicit posterior updates.
  • Because the posterior treats arms as locally independent, the harness may systematically under-explore combinations of arms (e.g., new loss plus new architecture together); the paper does not test whether such cross-arm synergy matters.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. RecHarness is an agentic harness for recommender-model optimization that separates edit-direction selection from concrete code mutation. A Thompson-sampling bandit router selects among a human-defined set of edit arms using binary validation-derived rewards, while an LLM generates concrete hypotheses and code edits within the selected arm. A jump-basin mechanism activates higher-level structural arms when local improvements stagnate. The paper reports offline experiments on five Amazon sequential-recommendation models and three KuaiRec watch-time/ranking models, plus a 7-day online A/B test on a short-video advertising platform, and claims more stable and more budget-efficient optimization than LLM-reasoning search.

Significance. If the central claim holds, RecHarness is a practically useful contribution to automated recommender-system optimization. The decoupling of routing from code generation is a clean, well-motivated idea, and the paper backs it with an unusually broad evaluation: eight model families, two recommendation tasks, a public code release, and an industrial online A/B test. The offline gains are consistent across models, and the online result is directionally meaningful. The main limitation is that the ablation designed to isolate the routing mechanism is confounded with the human-authored arm set, and the paper provides no uncertainty quantification for any of its comparisons, despite 'stability' being a headline claim. These issues are fixable with additional control conditions and error bars, so the contribution is defensible in principle but needs revision.

major comments (3)
  1. [Ablation Study, Table 5 and Eq. (4)/(16)] The headline comparison against 'LLM-reasoning search' uses the w/o Bandit condition, which removes both Thompson routing and the predefined arm set A = {a1,...,aK}. RecHarness retains the human-authored arms, including jump arms activated by Eq. (16). Because the largest gain in Table 5 (24.00% vs 10.16%) is attributed to a jump arm, the comparison conflates the routing mechanism with the human-engineered search space. The TR w/ LLM condition holds the arm set fixed and isolates the routing rule, but it is not a substitute for a control in which the LLM is given free rein plus the explicit jump-basin scaffolding. Without such a control, the abstract's claim that RecHarness is 'more stable and budget-efficient than LLM-reasoning search' over-attributes what may be the effect of the predefined arm set and jump mechanism. Please add either a condition that removes only the Thompson router
  2. [Analysis Experiments, Figure 2, Table 5; Online A/B Test, Table 6] No error bars, confidence intervals, or significance tests appear anywhere in the paper. The 'stability' claim is supported only by point estimates; Table 3 reports averages over three runs without variance, Figure 2 shows trajectories without intervals, Table 5 reports trial-level ratios without any measure of uncertainty, and Table 6 gives online metrics as point percentages with no CI, p-value, or guardrail description. Since the central claim is explicitly about variance ('more stable') and budget efficiency, the absence of uncertainty quantification is load-bearing. Please report at least bootstrap confidence intervals over runs for the offline comparisons and include a CI or significance test for the online A/B metrics.
  3. [Overall Performance, Table 3 and Table 4 ('Paper' columns)] The comparison to Kim et al. (2025) and Ma et al. (2026) is presented as if it were a matched comparison, but the paper does not state whether the 'Paper' baselines use the same GPU-time budget (43,200s), the same validation-based promotion rule, or the same number of trials. Without matched protocols, these columns cannot support the claim that RecHarness's advantage comes from its routing rather than from additional compute, different model tuning, or a different evaluation protocol. Either match the protocol and state the matching explicitly, or relabel these columns as 'reported values' and avoid comparative claims based on them.
minor comments (6)
  1. [Methodology, Eq. (9) and Eq. (13)] The values of the Beta prior hyperparameters (α0, β0), group size G, jump window W, saturation threshold τ, jump acceptance margin δjump, retuning rounds R, and the normalization constant ε in Eq. (9) are not specified. Please provide the exact values or a sensitivity analysis, since the algorithm's behavior depends on them.
  2. [Online A/B Test narrative] The statement that 'the router estimates that the search is nearing a local basin ceiling' is qualitative. Please report the actual improvement-rate value v_t from Eq. (15) and the threshold τ used, so that the reader can verify the jump was triggered by the mechanism rather than by manual intervention.
  3. [Table 2] The KuaiRec statistics columns are confusing ('Avg Inter./User' values such as '~1,746' and '~3,314' look like ranges or conditional counts). Please clarify what these numbers represent and relate them to the preprocessing described in the text.
  4. [Figure 2] The figure would be easier to read with labeled line markers or a clear color legend, and with variance bands or error bars if available.
  5. [Eq. (10)] The reward is defined as positive only if the candidate 'outperforms the average candidate in the same trial group' (Eq. 9). Make explicit that this is a relative success criterion and that an arm can be marked unsuccessful even when its absolute validation score improves, which may affect the posterior interpretation.
  6. [Contributions] The phrase 'the first optimization harness for self-evolving recommender systems' is stronger than the related-work section supports, given the cited systems such as EvoRec, NOVA, and AgentX. Consider softening the claim.

Circularity Check

0 steps flagged

No significant circularity: the central claim is evaluated on held-out test data and an online A/B test, and no prediction reduces to fitted inputs.

full rationale

RecHarness makes an empirical, systems-level claim. The bandit posterior (Eqs. 10–13) is updated from validation outcomes before future trials are allocated, while the final test-set metrics (Table 3/4) and online A/B metrics (Table 6) are held out and are not used in the posterior update or promotion rule. No fitted parameter is renamed as a prediction: the edit-arm set in Eq. 4 is a human-specified search input, and the gains of the selected candidate are measured externally on the online platform. The jump-basin mechanism (Eqs. 15–17) is a heuristic activation rule with externally observable threshold behavior, not an identity that forces the reported improvements. The ablation 'w/o Bandit' is confounded because it removes both the predefined arm set and the router, so it does not isolate the routing contribution; however, this is an experimental-attribution concern, not an equation-level reduction of the result to its inputs. Self-citations (e.g., BERT4Rec, D2Q, RecGOAT/Taiji) serve as background or baseline references and are not load-bearing for the central claim. Therefore no specific circular step can be identified and quoted.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The system introduces no new physical or theoretical entities; components like Experiment Skill and the jump-basin mechanism are software artifacts. The main free parameters are the Thompson-sampling priors and the jump-basin thresholds, none of which are specified with values or sensitivity analyses. The most important axiom is that the human-defined arm set actually contains the useful modification directions.

free parameters (7)
  • Beta prior hyperparameters (α0, β0)
    Initial success/failure counts for Thompson sampling are not specified; they determine initial exploration behavior.
  • Group size G (parallel trials per round) = 4 (ablation)
    The ablation uses four parallel trials per round; the effect of this choice on routing variance is not studied.
  • Jump window W
    Eq. (15) computes recent improvement rate over W rounds; W is not specified.
  • Saturation threshold τ
    Eq. (16) activates jump arms when the improvement rate falls below τ; no value or sensitivity analysis is reported.
  • Jump acceptance margin δjump
    Eq. (17) requires a retuned jump branch to beat the incumbent by δjump; the margin is not reported.
  • Retuning rounds R
    Maximum number of local retuning rounds for a jump candidate (Eq. 17) is not reported.
  • Normalization constant epsilon in Eq. (9) = small constant
    Used for numerical stability in group normalization; minor but unspecified.
axioms (4)
  • domain assumption Human-defined arm set A contains the useful modification directions.
    Eq. (4) defines A as predetermined; any bottleneck outside A is unreachable, and the online win required a pre-defined 'sequence_encoder_upgrade' jump arm.
  • domain assumption Validation metric correlates with the target business metric.
    Routing and promotion use only validation scores (Eqs. 8-10); the online A/B is the ultimate check, but the search assumes the offline signal is informative.
  • domain assumption LLM can generate executable code edits within each selected arm.
    Eq. (6) relies on the mutation operator producing valid code; invalid trials are filtered, but code-generation reliability limits the system.
  • ad hoc to paper Beta-Bernoulli Thompson sampling is a sensible posterior despite non-stationary incumbents.
    The paper states it does not assume fixed arm success probabilities, yet uses Beta posteriors (Eq. 11); the posterior is treated as local evidence rather than a statistically valid model.

pith-pipeline@v1.3.0-daily-deepseek · 13183 in / 11107 out tokens · 109913 ms · 2026-08-03T10:56:01.173718+00:00 · methodology

0 comments
read the original abstract

Optimizing modern recommender models still depends heavily on engineers manually iterating over architectural, objective, and training-strategy changes. While LLM-based agents can automate this trial-and-error process, allowing the LLM to both select modification directions and generate concrete hypotheses often leads to unstable search under limited experiment budgets. Inspired by the above challenge, we propose RecHarness, a Bandit-Routed Agentic Harness for automated recommender model optimization. RecHarness separates the optimization process into two steps: a bandit router selects the next modification direction according to historical validation feedback, while the LLM generates a concrete optimization hypothesis and executable code edit within the selected direction. To sustain long-horizon exploration, RecHarness uses a jump-basin mechanism to activate a structural-jump arm when local edits stagnate. Across multiple recommendation tasks, datasets, and model backbones, RecHarness achieves more stable performance improvements and uses limited trial budgets more effectively than LLM-reasoning search. During a 7-day online A/B test on a large-scale short-video advertising platform, the selected candidate improves ADVV by 2.084%, Revenue by 0.534%, and Exposure by 0.559%. Code is available at https://github.com/6lyc/RecHarness.

Figures

Figures reproduced from arXiv: 2607.29241 by Chi Lu, Haoran Ling, Jing Yao, Peng Jiang, Shuwen Kang, Wenjin Wu, Yuecheng Li, Zeyu Song.

Figure 1
Figure 1. Figure 1: Overview of RecHarness. RecHarness operates across three levels: [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Best-so-far validation HR@10 during the SASRec [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

86 extracted references · 13 linked inside Pith

  1. [1]

    Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

    Auto-WEKA: Combined selection and hyperparameter optimization of classification algorithms , author=. Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

  2. [2]

    Advances in neural information processing systems , volume=

    Efficient and robust automated machine learning , author=. Advances in neural information processing systems , volume=

  3. [3]

    ACM Transactions on Information Systems , volume=

    Automl for deep recommender systems: A survey , author=. ACM Transactions on Information Systems , volume=. 2023 , publisher=

    2023

  4. [4]

    Automated machine learning: Methods, systems, challenges , pages=

    Hyperparameter optimization , author=. Automated machine learning: Methods, systems, challenges , pages=. 2019 , publisher=

    2019

  5. [5]

    Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

    Optuna: A next-generation hyperparameter optimization framework , author=. Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

  6. [6]

    Advances in neural information processing systems , volume=

    Neural architecture optimization , author=. Advances in neural information processing systems , volume=

  7. [7]

    2019 , url=

    Hanxiao Liu and Karen Simonyan and Yiming Yang , booktitle=. 2019 , url=

    2019

  8. [8]

    ACM Computing Surveys (CSUR) , volume=

    A comprehensive survey of neural architecture search: Challenges and solutions , author=. ACM Computing Surveys (CSUR) , volume=. 2021 , publisher=

    2021

  9. [9]

    The Eleventh International Conference on Learning Representations , year=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. The Eleventh International Conference on Learning Representations , year=

  10. [10]

    International Conference on Learning Representations , volume=

    Mle-bench: Evaluating machine learning agents on machine learning engineering , author=. International Conference on Learning Representations , volume=

  11. [12]

    Advances in Neural Information Processing Systems , volume=

    Ai research agents for machine learning: Search, exploration, and generalization in mle-bench , author=. Advances in Neural Information Processing Systems , volume=

  12. [13]

    Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    Reasoning as Gradient: Scaling MLE Agents Beyond Tree Search , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    2026

  13. [14]

    Proceedings of the ACM Web Conference 2022 , pages=

    Autofield: Automating feature selection in deep recommender systems , author=. Proceedings of the ACM Web Conference 2022 , pages=

    2022

  14. [15]

    Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , pages=

    Autoloss: Automated loss function search in recommendations , author=. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , pages=

  15. [16]

    2022 IEEE 38th International Conference on Data Engineering (ICDE) , pages=

    Memorize, factorize, or be naive: Learning optimal feature interaction methods for CTR prediction , author=. 2022 IEEE 38th International Conference on Data Engineering (ICDE) , pages=. 2022 , organization=

    2022

  16. [22]

    Advances in Neural Information Processing Systems , volume=

    Can large language models explore in-context? , author=. Advances in Neural Information Processing Systems , volume=

  17. [23]

    arXiv preprint arXiv:2506.16499 , year=

    Ml-master: Towards ai-for-ai via integration of exploration and reasoning , author=. arXiv preprint arXiv:2506.16499 , year=

    Pith/arXiv arXiv

  18. [24]

    Advances in Neural Information Processing Systems , volume=

    Mle-star: Machine learning engineering agent via search and targeted refinement , author=. Advances in Neural Information Processing Systems , volume=

  19. [25]

    2018 IEEE international conference on data mining (ICDM) , pages=

    Self-attentive sequential recommendation , author=. 2018 IEEE international conference on data mining (ICDM) , pages=. 2018 , organization=

    2018

  20. [26]

    Proceedings of the 28th ACM international conference on information and knowledge management , pages=

    BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer , author=. Proceedings of the 28th ACM international conference on information and knowledge management , pages=

  21. [28]

    Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Tree based progressive regression model for watch-time prediction in short-video recommendation , author=. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  22. [29]

    Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining , pages=

    Deconfounding duration bias in watch-time prediction for video recommendation , author=. Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining , pages=

  23. [30]

    Proceedings of the ACM Web Conference 2026 , pages=

    Generative regression based watch time prediction for short-video recommendation , author=. Proceedings of the ACM Web Conference 2026 , pages=

    2026

  24. [31]

    arXiv preprint arXiv:2503.01067 , year=

    All roads lead to likelihood: The value of reinforcement learning in fine-tuning , author=. arXiv preprint arXiv:2503.01067 , year=

    arXiv

  25. [32]

    Biometrika , volume=

    On the likelihood that one unknown probability exceeds another in view of the evidence of two samples , author=. Biometrika , volume=. 1933 , publisher=

    1933

  26. [33]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Multi-armed bandits meet large language models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  27. [34]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Lost in Sequence: Do Large Language Models Understand Sequential Recommendation? , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2 , pages=

  28. [35]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Large language model-enhanced multi-armed bandits , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  29. [37]

    gradient descent

    Automatic prompt optimization with “gradient descent” and beam search , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  30. [38]

    Advances in neural information processing systems , volume=

    Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=

  31. [39]

    Advances in neural information processing systems , volume=

    Self-refine: Iterative refinement with self-feedback , author=. Advances in neural information processing systems , volume=

  32. [40]

    Nature , volume=

    Optimizing generative ai by backpropagating language model feedback , author=. Nature , volume=. 2025 , publisher=

    2025

  33. [43]

    Proceedings of the twelfth ACM international conference on web search and data mining , pages=

    A simple convolutional generative network for next item recommendation , author=. Proceedings of the twelfth ACM international conference on web search and data mining , pages=

  34. [45]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Bridging Language and Items for Retrieval and Recommendation: Benchmarking LLMs as Semantic Encoders , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  35. [46]

    Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages=

    KuaiRec: A fully-observed dataset and insights for evaluating recommender systems , author=. Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages=

  36. [47]

    Proceedings of the Nineteenth ACM Conference on Recommender Systems , pages=

    Longer: Scaling up long sequence modeling in industrial recommenders , author=. Proceedings of the Nineteenth ACM Conference on Recommender Systems , pages=

  37. [48]

    Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

    Lightgcn: Simplifying and powering graph convolution network for recommendation , author=. Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

  38. [49]

    Advances in Neural Information Processing Systems , volume=

    Structured Spectral Reasoning for Frequency-Adaptive Multimodal Recommendation , author=. Advances in Neural Information Processing Systems , volume=

  39. [52]

    Akiba, T.; Sano, S.; Yanase, T.; Ohta, T.; and Koyama, M. 2019. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2623--2631

    2019

  40. [53]

    Bouneffouf, D.; and Feraud, R. 2026. Multi-armed bandits meet large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 39682--39690

    2026

  41. [54]

    Chai, Z.; Ren, Q.; Xiao, X.; Yang, H.; Han, B.; Zhang, S.; Chen, D.; Lu, H.; Zhao, W.; Yu, L.; et al. 2025. Longer: Scaling up long sequence modeling in industrial recommenders. In Proceedings of the Nineteenth ACM Conference on Recommender Systems, 247--256

    2025

  42. [55]

    S.; Chowdhury, N.; Jaffe, O.; Aung, J.; Sherburn, D.; Mays, E.; Starace, G.; Liu, K.; Maksin, L.; Patwardhan, T.; et al

    Chan, J. S.; Chowdhury, N.; Jaffe, O.; Aung, J.; Sherburn, D.; Mays, E.; Starace, G.; Liu, K.; Maksin, L.; Patwardhan, T.; et al. 2025. Mle-bench: Evaluating machine learning agents on machine learning engineering. In International Conference on Learning Representations, volume 2025, 50466--50494

    2025

  43. [56]

    Cheng, Y.; Zhou, L.; Liang, X.; Luo, D.; Lee, T.; Zheng, K.; Zhang, W.; Cai, M.; Dong, J.; and Zhang, A. 2026. Let the Agent Steer: Closed-Loop Ranking Optimization via Influence Exchange. arXiv preprint arXiv:2603.27765

    Pith/arXiv arXiv 2026

  44. [57]

    Feurer, M.; and Hutter, F. 2019. Hyperparameter optimization. In Automated machine learning: Methods, systems, challenges, 3--33. Springer

    2019

  45. [58]

    Feurer, M.; Klein, A.; Eggensperger, K.; Springenberg, J.; Blum, M.; and Hutter, F. 2015. Efficient and robust automated machine learning. Advances in neural information processing systems, 28

    2015

  46. [59]

    Gao, C.; Li, S.; Lei, W.; Chen, J.; Li, B.; Jiang, P.; He, X.; Mao, J.; and Chua, T.-S. 2022. KuaiRec: A fully-observed dataset and insights for evaluating recommender systems. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, 540--550

    2022

  47. [60]

    He, X.; Deng, K.; Wang, X.; Li, Y.; Zhang, Y.; and Wang, M. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, 639--648

    2020

  48. [61]

    Hidasi, B.; Karatzoglou, A.; Baltrunas, L.; and Tikk, D. 2015. Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939

    Pith/arXiv arXiv 2015

  49. [62]

    Hou, Y.; Li, J.; Fu, X.; He, Z.; Yan, A.; Chen, X.; and McAuley, J. 2026. Bridging Language and Items for Retrieval and Recommendation: Benchmarking LLMs as Semantic Encoders. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 3251--3265

    2026

  50. [63]

    Huang, Q.; Vora, J.; Liang, P.; and Leskovec, J. 2023. Mlagentbench: Evaluating language agents on machine learning experimentation. arXiv preprint arXiv:2310.03302

    Pith/arXiv arXiv 2023

  51. [64]

    Jiang, Z.; Schmidt, D.; Srikanth, D.; Xu, D.; Kaplan, I.; Jacenko, D.; and Wu, Y. 2025. Aide: Ai-driven exploration in the space of code. arXiv preprint arXiv:2502.13138

    Pith/arXiv arXiv 2025

  52. [65]

    Kang, W.-C.; and McAuley, J. 2018. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM), 197--206. IEEE

    2018

  53. [66]

    Kim, S.; Kang, H.; Kim, K.; Kim, J.; Kim, D.; Yang, M.; Oh, K.; McAuley, J.; and Park, C. 2025. Lost in Sequence: Do Large Language Models Understand Sequential Recommendation? In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, 1160--1171

    2025

  54. [67]

    J.; Zhang, C.; and Slivkins, A

    Krishnamurthy, A.; Harris, K.; Foster, D. J.; Zhang, C.; and Slivkins, A. 2024. Can large language models explore in-context? Advances in Neural Information Processing Systems, 37: 120124--120158

    2024

  55. [68]

    Lao, C.; Pan, F.; Ma, G.; Li, H.; Lin, H.; Shi, J.; Zhao, K.; Gai, K.; Zhou, M.; Zhou, Q.; et al. 2026. AgentX: Towards Agent-Driven Self-Iteration of Industrial Recommender Systems. arXiv preprint arXiv:2606.26859

    Pith/arXiv arXiv 2026

  56. [69]

    Li, Y.; Ju, H.; Song, Z.; Yang, W.; Lu, C.; Jiang, P.; and Gai, K. 2026 a . RecGOAT: Graph Optimal Adaptive Transport for LLM-Enhanced Multimodal Recommendation with Dual Semantic Alignment. arXiv preprint arXiv:2602.00682

    Pith/arXiv arXiv 2026

  57. [70]

    Li, Y.; Song, Z.; Yao, J.; Lu, C.; Jiang, P.; and Gai, K. 2026 b . Taiji: Pareto Optimal Policy Optimization with Semantics-IDs Trade-off for Industrial LLM-Enhanced Recommendation. arXiv preprint arXiv:2606.03866

    Pith/arXiv arXiv 2026

  58. [71]

    Lin, X.; Chen, X.; Song, L.; Liu, J.; Li, B.; and Jiang, P. 2023. Tree based progressive regression model for watch-time prediction in short-video recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 4497--4506

    2023

  59. [72]

    Liu, H.; Simonyan, K.; and Yang, Y. 2019. DARTS : Differentiable Architecture Search. In International Conference on Learning Representations

    2019

  60. [73]

    Liu, S.; Fang, L.; Sun, Y.; Huang, S.; Luo, Q.; Chen, X.; Liu, D.; Ma, C.; Chai, Z.; Wang, H.; et al. 2026. NOVA: A Verification-Aware Agent Harness for Architecture Evolution in Industrial Recommender Systems. arXiv preprint arXiv:2606.27243

    Pith/arXiv arXiv 2026

  61. [74]

    Luo, R.; Tian, F.; Qin, T.; Chen, E.; and Liu, T.-Y. 2018. Neural architecture optimization. Advances in neural information processing systems, 31

    2018

  62. [75]

    Lyu, F.; Tang, X.; Guo, H.; Tang, R.; He, X.; Zhang, R.; and Liu, X. 2022. Memorize, factorize, or be naive: Learning optimal feature interaction methods for CTR prediction. In 2022 IEEE 38th International Conference on Data Engineering (ICDE), 1450--1462. IEEE

    2022

  63. [76]

    Ma, H.; Tian, K.; Zhang, T.; Zhang, X.; Zhou, H.; Jin, C.; Chen, C.; Li, H.; Guan, J.; and Zhou, S. 2026. Generative regression based watch time prediction for short-video recommendation. In Proceedings of the ACM Web Conference 2026, 6183--6193

    2026

  64. [77]

    Madaan, A.; Tandon, N.; Gupta, P.; Hallinan, S.; Gao, L.; Wiegreffe, S.; Alon, U.; Dziri, N.; Prabhumoye, S.; Yang, Y.; et al. 2023. Self-refine: Iterative refinement with self-feedback. Advances in neural information processing systems, 36: 46534--46594

    2023

  65. [78]

    Mu, L.; Deng, H.; Xing, H.; Hu, J.; Zhang, Y.; and Zeng, X. 2026. EvoRec: Self Evolving Agentic Recommender Systems. arXiv preprint arXiv:2606.28368

    Pith/arXiv arXiv 2026

  66. [79]

    Nam, J.; Yoon, J.; Chen, J.; Shin, J.; Arik, S.; and Pfister, T. 2026. Mle-star: Machine learning engineering agent via search and targeted refinement. Advances in Neural Information Processing Systems, 38: 116692--116712

    2026

  67. [80]

    X.; Li, W.; Zhang, L.; Chen, S.; and Wen, J.-R

    Ou, K.; Wu, C.; Wang, X.; Zheng, B.; Zhao, W. X.; Li, W.; Zhang, L.; Chen, S.; and Wen, J.-R. 2026. Deep Research for Recommender Systems. arXiv preprint arXiv:2603.07605

    arXiv 2026

  68. [81]

    gradient descent

    Pryzant, R.; Iter, D.; Li, J.; Lee, Y.; Zhu, C.; and Zeng, M. 2023. Automatic prompt optimization with “gradient descent” and beam search. In Proceedings of the 2023 conference on empirical methods in natural language processing, 7957--7968

    2023

  69. [82]

    Ren, P.; Xiao, Y.; Chang, X.; Huang, P.-Y.; Li, Z.; Chen, X.; and Wang, X. 2021. A comprehensive survey of neural architecture search: Challenges and solutions. ACM Computing Surveys (CSUR), 54(4): 1--34

    2021

  70. [83]

    Shinn, N.; Cassano, F.; Gopinath, A.; Narasimhan, K.; and Yao, S. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in neural information processing systems, 36: 8634--8652

    2023

  71. [84]

    Sun, F.; Liu, J.; Wu, J.; Pei, C.; Lin, X.; Ou, W.; and Jiang, P. 2019. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management, 1441--1450

    2019

  72. [85]

    C.; and Dai, Z

    Sun, J.; Wang, Z.; Yang, R.; Xiao, C.; Lui, J. C.; and Dai, Z. 2026. Large language model-enhanced multi-armed bandits. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 8130--8145

    2026

  73. [86]

    H.; and Leyton-Brown, K

    Thornton, C.; Hutter, F.; Hoos, H. H.; and Leyton-Brown, K. 2013. Auto-WEKA: Combined selection and hyperparameter optimization of classification algorithms. In Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining, 847--855

    2013

  74. [87]

    Toledo, E.; Hambardzumyan, K.; Josifoski, M.; Hazra, R.; Baldwin, N.; Audran-Reiss, A.; Kuchnik, M.; Magka, D.; Jiang, M.; Lupidi, A.; et al. 2026. Ai research agents for machine learning: Search, exploration, and generalization in mle-bench. Advances in Neural Information Processing Systems, 38: 35309--35348

    2026

  75. [88]

    Wang, H.; Wu, Y.; Chang, D.; Wei, L.; and Heldt, L. 2026. Self-evolving recommendation system: End-to-end autonomous model optimization with LLM agents. arXiv preprint arXiv:2602.10226

    Pith/arXiv arXiv 2026

  76. [89]

    Wang, Y.; Zhao, X.; Xu, T.; and Wu, X. 2022. Autofield: Automating feature selection in deep recommender systems. In Proceedings of the ACM Web Conference 2022, 1977--1986

    2022

  77. [90]

    Yang, W.; Zhong, R.; Chen, Y.; Lu, C.; and Jiang, P. 2026. Structured Spectral Reasoning for Frequency-Adaptive Multimodal Recommendation. Advances in Neural Information Processing Systems, 38: 28122--28143

    2026

  78. [91]

    R.; and Cao, Y

    Yao, S.; Zhao, J.; Yu, D.; Du, N.; Shafran, I.; Narasimhan, K. R.; and Cao, Y. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In The Eleventh International Conference on Learning Representations

    2023

  79. [92]

    M.; and He, X

    Yuan, F.; Karatzoglou, A.; Arapakis, I.; Jose, J. M.; and He, X. 2019. A simple convolutional generative network for next item recommendation. In Proceedings of the twelfth ACM international conference on web search and data mining, 582--590

    2019

  80. [93]

    Yuksekgonul, M.; Bianchi, F.; Boen, J.; Liu, S.; Lu, P.; Huang, Z.; Guestrin, C.; and Zou, J. 2025. Optimizing generative ai by backpropagating language model feedback. Nature, 639(8055): 609--616

    2025

Showing first 80 references.