Shopping Companion: Benchmarking and Training LLM Agents for Long-Horizon Preference-Grounded E-Commerce Tasks

Huaipeng Zhao; Kejun Xiao; Tao Luo; Xiaoyi Zeng; Zijian Yu

REVIEW 2 major objections 1 minor 3 cited by

Shopping Companion: Benchmarking and Training LLM Agents for Long-Horizon Preference-Grounded E-Commerce Tasks

T0 review · 2 major / 1 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Annotation-free tool-wise rewards let a fine-tuned 4B model outperform strong baselines on long-horizon preference-grounded shopping, where even frontier models stay under 70% success.

desk verdict We only have the Shopping Companion abstract; the attached full text is a different paper (EnSF data assimilation), so the shopping claims cannot be audited. read the letter →

arxiv 2603.14864 v3 pith:L6C4LIYU submitted 2026-03-16 cs.CL

Zijian Yu , Kejun Xiao , Huaipeng Zhao , Tao Luo , Xiaoyi Zeng This is my paper · ORCID

classification cs.CL

keywords LLMagentse-commercepreferencememoryprocesssupervisiontool-wiserewardsshoppingbenchmarklong-horizontasks

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

LLM shopping agents must track user preferences across long, multi-session conversations, yet the field has lacked both hard evaluation suites and dense training signals. This paper introduces Shopping Companion Bench: two tasks that require cross-session preference memory over a pool of more than 1.2 million real products. Failure analysis isolates two main error sources—cascading preference hallucination and weak verification of product attributes against stated requirements. To fix them, the authors design annotation-free rewards that score each tool call, giving process supervision instead of a single sparse terminal signal. A lightweight 4B model trained with these rewards consistently beats strong baselines on preference capture and task success, while even state-of-the-art models remain below 70% success, showing both the benchmark’s difficulty and the reward design’s value.

What carries the argument

Shopping Companion Bench plus annotation-free tool-wise rewards that assign process supervision to every tool call; these rewards specifically target preference hallucination and attribute-verification failures so credit is no longer delayed until task end.

What would settle it

An independent evaluation with real human shoppers or a held-out product pool in which the 4B model no longer beats baselines, or in which tool-wise rewards fail to reduce preference hallucination relative to terminal-only rewards, would falsify the central claim.

Watch

Extended reading notes

Core claim

Annotation-free, tool-wise process rewards alleviate reward sparsity on long-horizon preference-grounded shopping tasks, enabling a fine-tuned 4B model to outperform strong baselines on Shopping Companion Bench—a new two-task suite requiring cross-session preference memory over 1.2 million real items—where even models such as GPT-5 stay below 70% success.

Load-bearing premise

That success on the two constructed tasks under the authors’ tool-wise rewards measures real long-horizon shopping competence rather than artifacts of synthetic user trajectories, product-pool construction, or reward hacking of the tool interface.

Editorial extensions

If this is right

Long-horizon shopping agents can be trained without expensive human process annotations.
Lightweight models become competitive for preference-grounded e-commerce once dense tool-level rewards are available.
Preference hallucination and attribute verification become first-class failure modes that future agents must explicitly address.
Cross-session memory benchmarks over large real catalogs will stay hard for frontier models until process supervision improves.
Tool-level reward design can transfer to other multi-step agent domains that suffer from sparse terminal rewards.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same tool-wise process rewards could improve agents for multi-session preference tasks outside shopping, such as travel planning or medical intake.
Larger models may close or reverse the gap if they are also trained with the same dense tool rewards rather than only terminal success.
Reported gains will largely depend on how faithfully the synthetic user trajectories match real shopper behavior.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Punchline first: the manuscript body in the cache is not Shopping Companion. It is a score-filter / EnSF data-assimilation paper on Lorenz-96 and KdV. For arXiv:2603.14864 we effectively have the abstract only. Treat every number below as unverified.

From that abstract alone, the intended contribution is clear and useful within agents/e-commerce: a long-horizon preference-memory shopping benchmark on a large real catalog (~1.2M items), a failure-mode diagnosis (preference hallucination cascading, weak attribute checks), and annotation-free tool-wise process rewards aimed at sparse long-horizon credit. Claiming GPT-class models stay under 70% and that a fine-tuned 4B beats strong baselines is the kind of result people in the area would actually use if it holds.

What I cannot check: task construction and synthetic-user validity, whether “success” is gamed by the tool interface, ablations that separate the reward design from the benchmark, baselines, variance, and any leakage between reward signals and evaluation. The circularity risk is structural—same group defines tasks, rewards, and metrics—not proof of bad faith, but it is the main soft spot until independent reimplementation or external validation exists.

The EnSF body that was supplied is a different, serious-looking dynamical-systems paper; it does not support or refute the shopping claims. So this is not a soft spot in Shopping Companion’s math; it is a packaging/mismatch problem that blocks review of the titled work.

Who it is for: people building shopping or long-horizon tool agents who need harder preference-memory evals. Value is real if the full paper and artifacts match the abstract. I would send a complete, consistent manuscript to peer review. I would not cite or run a reading group on the abstract alone. Get the correct PDF and code before spending more time.

Referee Report

2 major / 1 minor

Summary. The submission is presented under the title and abstract of “Shopping Companion: Benchmarking and Training LLM Agents for Long-Horizon Preference-Grounded E-Commerce Tasks” (arXiv:2603.14864, cs.CL). That abstract claims a new long-horizon preference-grounded shopping benchmark (Shopping Companion Bench) over a 1.2M-item product pool, a failure-mode analysis (preference hallucination and attribute verification), annotation-free tool-wise process rewards, and experimental results in which GPT-5 stays below 70% success while a fine-tuned 4B model outperforms strong baselines. The body that was supplied, however, is an entirely different manuscript: “A Score Filter Enhanced Data Assimilation Framework for Data-Driven Dynamical Systems” (content matching arXiv:2603.14863), which develops an Ensemble Score Filter (EnSF) hybrid with LSTM and R-DeepONet surrogates and evaluates it on Lorenz–96 and the KdV equation. No Shopping Companion methods, tasks, rewards, or results appear in the body.

Significance. If the abstract’s claims were supported by a matching manuscript, a hard long-horizon preference-memory shopping benchmark with process-level, annotation-free tool rewards and a competitive lightweight agent would be of clear interest to the LLM-agent and e-commerce communities. As submitted, that contribution cannot be assessed: the body contains no benchmark definition, no agent training pipeline, and no e-commerce experiments. The EnSF–ML assimilation work in the body is a separate, technically coherent line of research on uncertainty reduction for data-driven dynamical systems, but it is not the paper announced by the title and abstract. Significance of the claimed Shopping Companion result is therefore indeterminate from this package.

major comments (2)

[Title, Abstract, and full body (Sections 1–5)] Title/abstract vs. body mismatch: the abstract and paper_id announce Shopping Companion Bench, tool-wise rewards, GPT-5 <70% success, and a fine-tuned 4B agent, but Sections 1–5 and all figures/tables develop EnSF-enhanced LSTM/R-DeepONet forecasting on Lorenz–96 and KdV. There is no section defining the two shopping tasks, the 1.2M product pool, preference memory, tool-wise rewards, or any LLM-agent experiment. The central claims of 2603.14864 are therefore unsupported by the supplied manuscript and cannot be audited for soundness, baselines, or ablations.
[Abstract claims vs. Sections 4.1–4.2 (which report only Lorenz–96/KdV RMSE)] Because the body is a different paper, load-bearing Shopping Companion claims (preference hallucination as a cascading failure mode; annotation-free tool-wise rewards alleviating reward sparsity; 4B outperformance of strong baselines; GPT-5 success below 70%) have no methods, equations, tables, or error bars against which to check them. Evaluation of those claims is not possible without the correct manuscript body.

minor comments (1)

[Section 2.1, Eq. (1); Figures 3–15] The supplied body (EnSF paper) itself has ordinary presentation issues—e.g., garbled math glyphs in the LSTM equations (Section 2.1), inconsistent notation for observation operators, and figure captions that sometimes restate results without quantitative detail—but these are secondary to the abstract/body identity failure and would only matter if that paper were the submission under review.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EnSF+ML forecasts are evaluated against external numerical ground truth (Lorenz-96, KdV) and EnKF baselines; self-citations supply the filter method, not a tautological result.

full rationale

The supplied full manuscript is the EnSF data-assimilation paper (arXiv:2603.14863), not the Shopping Companion abstract (2603.14864). On the manuscript that is actually present: the load-bearing chain is (i) train LSTM / R-DeepONet surrogates, (ii) roll them out in long-term prediction mode where errors accumulate, (iii) correct the forecast ensemble with EnSF (or EnKF) using noisy/partial observations, (iv) report RMSE against trajectories from classical numerical solvers. That chain is not circular. EnSF is imported from the authors’ prior work as a filtering method, but the claimed improvement is an empirical comparison to EnKF and to uncorrected LTP/SSP on external dynamical systems, not a quantity forced by definition or by a uniqueness theorem. Score approximations (mini-batch score, likelihood injection with damping h(t)=1−t) are algorithmic choices, not fits renamed as predictions. No self-definitional loop, no fitted-input-as-prediction, and no load-bearing uniqueness import appear in the derivation. Honest non-finding: score 0; steps empty.

Assumptions & free parameters 0 free parameters · 3 assumptions · 2 invented entities

Abstract-only review of Shopping Companion Bench. Load-bearing premises are domain assumptions about what constitutes long-horizon preference-grounded shopping competence, that tool-call process rewards without annotation are valid supervision, and that the constructed product pool and tasks are representative. No free parameters or invented physical entities are stated in the abstract; the main invented construct is the benchmark/reward package itself.

assumptions (3)

domain assumption Cross-session preference memory and attribute verification against user requirements are the critical bottlenecks for e-commerce LLM agents on long-horizon tasks.
Abstract frames these as the two major failure sources motivating the benchmark and rewards.
ad hoc to paper Annotation-free tool-wise rewards provide sufficient process supervision to alleviate reward sparsity without human step labels.
Central training claim; abstract asserts effectiveness via 4B fine-tune gains but does not derive necessity or optimality.
domain assumption Success rates on the two Shopping Companion Bench tasks are a meaningful proxy for real preference-grounded shopping performance over large catalogs.
Benchmark validity assumption; required for interpreting GPT-5 <70% and 4B outperformance as field progress.

invented entities (2)

Shopping Companion Bench
purpose: Provide evaluation tasks requiring cross-session preference memory over a 1.2M+ real-product pool.
Primary contribution named in the abstract; independent evidence outside this paper is not shown here.
Annotation-free tool-wise rewards
purpose: Supply dense process supervision per tool call to train shopping agents without step annotations.
Methodological construct claimed to fix preference hallucination and weak attribute verification; formulas not given in abstract.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Shopping Companion: Benchmarking and Training LLM Agents for Long-Horizon Preference-Grounded E-Commerce Tasks." pith.science (2026). https://pith.science/paper/L6C4LIYU

@misc{pith2026260314864,
  author       = {Pith},
  title        = {Pith review of: Shopping Companion: Benchmarking and Training LLM Agents for Long-Horizon Preference-Grounded E-Commerce Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L6C4LIYU}},
  note         = {Machine review of arXiv:2603.14864}
}

read the original abstract

In e-commerce, LLM agents show promise for shopping tasks such as recommendations, budget management, and bundle deals, where accurately capturing user preferences from long-horizon conversations is critical. However, progress is limited by two key challenges: (1) the absence of benchmarks for evaluating long-term preference-aware shopping tasks, and (2) the lack of fine-grained supervision for shopping agent training. To fill the benchmark gap, we introduce Shopping Companion Bench, a novel benchmark comprising two shopping tasks that require cross-session preference memory, grounded in a product pool of over 1.2 million real-world items. Our analysis further identifies two major sources of failure on this benchmark: cascading errors caused by preference hallucination, and insufficient verification of product attributes against user requirements. To address these failure modes, we design annotation-free, tool-wise rewards that provide process supervision for each tool call, alleviating reward sparsity in long-horizon tasks. Experimental results demonstrate that even state-of-the-art models such as GPT-5 achieve success rates below 70%, highlighting the difficulty of our benchmark. Notably, our fine-tuned lightweight 4B model consistently outperforms strong baselines in both preference capture and task performance, suggesting the effectiveness of our reward design.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

MMShopBench: A Real-Log Benchmark for Multimodal, Multi-Turn Shopping Agents
cs.AI 2026-07 conditional novelty 7.0 of 10

MMShopBench, built from real multimodal shopping logs, shows even the best agent satisfies the full request in under two-thirds of cases, and fine-tuning on 900 real-log trajectories substantially closes the gap to pr...
Can LLM Agents Price Competitively? A Dynamic Multi-Attribute Auction Benchmark for Agentic Commerce
cs.AI 2026-07 conditional novelty 7.0 of 10

LLM merchant agents in a new dynamic auction benchmark capture at most 32% of hindsight-optimal profit; profit tracks margin per win more than win rate, and fast pre-shock learners adapt poorly to preference shocks.
Paying to Know: Micro-Transaction Markets for Verified Product Information in Agentic E-Commerce
cs.CL 2026-06 unverdicted novelty 5.0 of 10

Agentic e-commerce should operate as a micro-transaction market for verified information unlocked progressively by buyer agents, redirecting NLP research toward cost-optimal acquisition, data pricing, and related problems.

Reference graph

Works this paper leans on

88 extracted references · 3 linked inside Pith · cited by 3 Pith papers

[1]

Allen, S

Anna A. Allen, S. Markou, W. Tebbutt, J. Requeima, W. P. Br uinsma, T. R. Andersson, M. Herzog, N. D. Lane, M .Chantry, J. S. Hosking, and R. E. Turn er. End-to-end data-driven weather prediction. ��, 641(8065):1172–1179, 5 2025. 21

2025
[2]

Andrieu, A

C. Andrieu, A. Doucet, and R. Holenstein. Particle marko v chain monte carlo methods. �� , 72(3):269–342, 2010

2010
[3]

Archibald and F

R. Archibald and F. Bao. A kernel learning method for back ward sde ﬁlter. �� , 2022

2022
[4]

Archibald, F

R. Archibald, F. Bao, and X. Tu. A direct method for parame ter estimations. �� , 398:108871, 2019

2019
[5]

F. Bao, Y. Cao, and X. Han. Forward backward doubly stocha stic diﬀerential equations and optimal ﬁltering of diﬀusion processes. �� , 18(3):635– 661, 2020

2020
[6]

F. Bao, Y. Cao, and J. Yong. Data informed solution estima tion for forward backward stochastic diﬀerential equations. �� , 19:439–464, 2021

2021
[7]

F. Bao, Y. Cao, and W. Zhao. A backward doubly stochastic d iﬀerential equation approach for nonlinear ﬁltering problems. �� , 23:1573–1601, 201
[8]

F. Bao, H. G. Chipilski, S. Liang, G. Zhang, and J. S. Whita ker. Nonlinear ensemble ﬁltering with diﬀusion models: Application to the surface quasigeost rophic dynamics. �� , 153(7):1155–1169, 2025

2025

Show all 88 references

[9]

F. Bao, Z. Zhang, and G. Zhang. An ensemble score ﬁlter for tracking high-dimensional nonlinear dynamical systems. �� , 432:117447, 2024

2024
[10]

F. Bao, Z. Zhang, and G. Zhang. A score-based ﬁlter for no nlinear data assimilation. �� , 514:113207, 2024

2024
[11]

United ﬁlter for jointly esti mating state and parameters of stochastic dynamical systems

F Bao, Z Zhang, and G Zhang. United ﬁlter for jointly esti mating state and parameters of stochastic dynamical systems. ��, 39(3):747–774, 2026

2026
[12]

Beneˇ s, M

K. Beneˇ s, M. K. Baskar, and L. Burget. Residual memory n etworks in language modeling: Improving the reputation of feed-forward networks. �� , pages 284–288, 2017

2017
[13]

Beucler, M

T. Beucler, M. Pritchard, S. Rasp, J. Ott, P. Baldi, and P . Gentine. Enforcing analytic con- straints in neural networks emulating physical systems. �� , 126:098302, 2021

2021
[14]

M. Buehner. Local ensemble transform kalman ﬁlter with cross validation. �� , 148:2265– 2282, 2020

2020
[15]

� ��

Michael C Burkhart. � �� . ProQuest Dissertations Publishing, 2019

2019
[16]

Chai and A

J. Chai and A. Li. Deep learning in natural language proc essing: A state-of-the-art survey. In �� , pages 1–6, 2019

2019
[17]

J. Chai, H. Zeng, A. Li, and E. W.T. Ngai. Deep learning in computer vision: A critical review of emerging techniques and application scenarios. �� , 6:100134, 2021. 22

2021
[18]

On the brownian displacements and ther mal diﬀusion of grains suspended in a non-uniform ﬂuid

Sydney Chapman. On the brownian displacements and ther mal diﬀusion of grains suspended in a non-uniform ﬂuid. �� , pages 34–54, 1928

1928
[19]

Chattopadhyay, P

A. Chattopadhyay, P. Hassanzadeh, and D. Subramanian. Data-driven predictions of a multi- scale lorenz 96 chaotic system using machine-learning meth ods: reservoir computing, artiﬁcial neural network, and long short-term memory network. �� , 27(3)...

2020
[20]

Chen and H

T. Chen and H. Chen. Universal approximation to nonline ar operators by neural networks with arbitrary activation functions and its application to dyna mical systems. �� , 6(4):911–917, 1995

1995
[21]

Chung, C

J. Chung, C. Gulcehre, K. Cho, and Y. Yere. Empirical eva luation of gated recurrent neural networks on sequence modeling. 2014

2014
[22]

An introduction to estimation theory (g tspecial issueltdata assimilation in meteology and oceanography: theory and practice)

Stephen E Cohn. An introduction to estimation theory (g tspecial issueltdata assimilation in meteology and oceanography: theory and practice). �� , 75(1B):257–288, 1997

1997
[23]

G. Evensen. The ensemble kalman ﬁlter: Theoretical for mulation and practical implementation. �� , 53(4):343–367, 2003

2003
[24]

Sequential data assimilation with a nonl inear quasi-geostrophic model using monte carlo methods to forecast error statistics

Geir Evensen. Sequential data assimilation with a nonl inear quasi-geostrophic model using monte carlo methods to forecast error statistics. �� , 99(C5):10143–10162, 1994

1994
[25]

��

Gier Evensen. �� . Springer Science & Business Media, 2009

2009
[26]

F. A. Gers, J. Schmidhuber, and F. Cummins. Learning to f orget: continual prediction with lstm. �� , 12(8):2451–2471, 2000

2000
[27]

Learning to forget: Continual prediction with lstm

Felix A Gers, J¨ urgen Schmidhuber, and Fred Cummins. Learning to forget: Continual prediction with lstm. �� , 12(10):2451–2471, 2000

2000
[28]

Goodfellow, Y

I. Goodfellow, Y. Bengio, and A. Courville. �� . MIT Press, 2016. deeplearning- book.org

2016
[29]

N. J. Gordon, D. J. Salmond, and A. F. M. Smith. Novel appr oach to nonlinear/non-gaussian bayesian state estimation. �� , 140(2):107–113, 1993

1993
[30]

Long short-term memory

Alex Graves. Long short-term memory. �� , pages 37–45, 2012

2012
[31]

��

Alex Graves. �� . Studies in Com- putational Intelligence. Springer, Berlin, Heidelberg, 2 012
[32]

Generating sequences with recurrent neur al networks

Alex Graves. Generating sequences with recurrent neur al networks. 2013

2013
[33]

K. Greﬀ, R. K. Srivastava, J. Koutn ´ ık, B. R. Steunebrink, and J. Schmidhuber. Lstm: A search space odyssey. �� , 28(10):2222–2232, 2016. 23

2016
[34]

G. J. Hakim and S. Masanam. Dynamical tests of a deep lear ning weather prediction model. �� , 3(3):e230090, 2024

2024
[35]

Z. Hao, J. Yao, C. Su, H. Su, Z. Wang, F. Lu, Z. Xia, Y. Zhang , S. Liu, Lu Lu, and J. Zhu. Pinnacle: A comprehensive benchmark of physics-informed n eural networks for solving pdes. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. To mczak, and C. Zhang, editors...

2024
[36]

Sequential deep operator networks (s-deeponet) for predic ting full-ﬁeld solutions under time- dependent loads

Junyan He, Shashank Kushwaha, Jaewan Park, Seid Koric, Diab Abueidda, and Iwona Jasiuk. Sequential deep operator networks (s-deeponet) for predic ting full-ﬁeld solutions under time- dependent loads. �� , 127:107258, 2024

2024
[37]

Hochreiter and J

S. Hochreiter and J. Schmidhuber. Long short-term memo ry. �� , 9(8):1735– 1780, 1997

1997
[38]

Long short-t erm memory

Sepp Hochreiter and J¨ urgen Schmidhuber. Long short-t erm memory. �� , 9(8):1735–1780, 1997

1997
[39]

Ensemble ka lman ﬁltering

Peter L Houtekamer and Herschel L Mitchell. Ensemble ka lman ﬁltering. �� , 131(613):3269–3289, 2005

2005
[40]

W. W. Hsieh and B. Tang. Applying neural network models t o prediction and data analysis in meteorology and oceanography. �� , 79(9):1855– 1870, 1998

1998
[41]

Hu and P

C.-C. Hu and P. J. van Leeuwen. A particle ﬂow ﬁlter for hi gh-dimensional system applications. �� , 147:2352–2374, 2021

2021
[42]

B. R. Hunt, E. J. Kostelich, and I. Szunyogh. Eﬃcient dat a assimilation for spatiotemporal chaos: A local ensemble transform kalman ﬁlter. �� , 230(1- 2):112–126, 2007

2007
[43]

Kolen J and S

F. Kolen J and S. C. Kremer. �� , pages 237–243. IEEE, 2001

2001
[44]

A. Jiao, H. He, R. Ranade, J. Pathak, and Lu Lu. One-shot l earning for solution operators of partial diﬀerential equations. �� , 16:8386, 2025

2025
[45]

R. E. Kalman. A new approach to linear ﬁltering and predi ction problems. �� , 82(1):35–45, 1960

1960
[46]

K. Kang, V. Maroulas, I. Schizas, and F. Bao. Improved di stributed particle ﬁlters for tracking in a wireless sensor network. �� , 117:90–108, 2018

2018
[47]

D. P. Kingma and J. Ba. Adam: A method for stochastic opti mization. In �� , Banﬀ, 14-16 April 2014

2014
[48]

¨Uber die analytischen methoden in der wahrscheinlichkeits rechnung

Andrei Kolmogoroﬀ. ¨Uber die analytischen methoden in der wahrscheinlichkeits rechnung. �� , pages 415 – 458, 1931. 24

1931
[49]

N. B. Kovachki, S. Lanthaler, and A. M. Stuart. Operator learning: Algorithms and analysis. arXiv preprint, arXiv:2402.15715, 2024

2024 arXiv
[50]

Tutorial: The kalman ﬁlter

Tony Lacey. Tutorial: The kalman ﬁlter. �� , 139, 1998

1998
[51]

Data ass imilation

Kody Law, Andrew Stuart, and Kostas Zygalakis. Data ass imilation. �� , 214:52, 2015

2015
[52]

Le and M

V.-A. Le and M. Dik. A mathematical analysis of neural op erator behaviors. arXiv preprint arXiv:2410.21481, 2024

2024
[53]

LeCun, Y

Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. ��, (521):436, 2015

2015
[54]

P. J. Van Leeuwen, H. R. K¨ unsch, L. Nerger, R. Potthast, and S. Reich. Particle ﬁlters for high- dimensional geoscience applications: A review. �� , 145(723):2335–2365, 2019

2019
[55]

Liang, H

S. Liang, H. Tran, F. Bao, H. G. Chipilski, P. Jan van Leeu wen, and G. Zhang. Ensemble score ﬁlter with image inpainting for data assimilation in tracki ng surface quasi-geostrophic dynamics with partial observations. ��
[56]

Z. C. Lipton, J. Berkowitz, and C. Elkan. A critical revi ew of recurrent neural networks for sequence learning. �� , 2015

2015
[57]

Y. Liu, M. Yang, Z. Zhang, Y. Cao, G. Zhang, and F. Bao. Diﬀu sion-model-assisted supervised learning of generative models for density estimation. �� , 5(1):25–38, 2024

2024
[58]

E. N. Lorenz. Predictability: A problem partly solved. In �� , volume 1, pages 1–18. Reading, 1996

1996
[59]

L. Lu, , G. Pang, P. Jin, Z. Zhang, and G. E. Karniadakis. L earning nonlinear operators via deeponet based on the universal approximation theorem o f operators. �� , 3:218–229, 2021

2021
[60]

L. Lu, P. Jin, and G. E. Karniadakis. Deeponet: Learning nonlinear operators for identifying diﬀerential equations based on the universal approximation theorem of operators. arXiv preprint, arXiv:1910.03193, 2019

1910 arXiv
[61]

L. Lu, X. Meng, S. Cai, Z. Mao, S. Goswami, Z. Zhang, and G. E. Karniadakis. A comprehensive and fair comparison of two neural operators (with practical extensions) based on fair data. �� , 393:114778, 2022

2022
[62]

Lu Lu, X. Meng, Z. Mao, and G. E. Karniadakis. Deepxde: A d eep learning library for solving diﬀerential equations. �� , 63(1):208–228, 2021

2021
[63]

S. Mao, A. M. Naser, S. N. Buoy, K. K. Brock, and K. A. Hutch eson. Optimizing modiﬁed barium swallow exam workﬂow: Automating pre-analysis vide o sorting in swallowing function assessment. In �� ...

2024
[64]

Neural operator learning for long-time integrat ion in dynamical systems with re- current neural networks

Katarzyna Micha/suppress lowska, Somdatta Goswami, George Em Karniadakis, and Signe Riemer- Sørensen. Neural operator learning for long-time integrat ion in dynamical systems with re- current neural networks. In �� , p...

2024
[65]

Morzfeld, X

M. Morzfeld, X. Tu, E. Atkins, and A. J. Chorin. A random m ap implementation of implicit ﬁlters. �� , 231(4):2049–2066, 2012
[66]

H. H. Nguyen and C. W. Chan. Multiple neural networks for a long term time series forecast. �� , 13(1):90–98, 2004

2004
[67]

D. W. Otter, J. R. Medina, and J. K. Kalita. A survey of the usages of deep learning for natural language processing. �� , 32(2):604–624, 2021

2021
[68]

Pascanu, T

R. Pascanu, T. Mikolov, and Y. Bengio. On the diﬃculty of training recurrent neural networks. In �� , pages 1310–1318. Pmlr, 2013

2013
[69]

Pathak, B

J. Pathak, B. Hunt, M. Girvan, Z. Lu, and E. Ott. Model-fr ee prediction of large spatiotempo- rally chaotic systems from data: A reservoir computing appr oach. �� , 120:024102, 1 2018

2018
[70]

M. K. Pitt and N. Shephard. Filtering via simulation: Au xiliary particle ﬁlters. �� , 94(446):590–599, 1999

1999
[71]

Radosavovic, R

I. Radosavovic, R. P. Kosaraju, R. Girshick, K. He, and P . Doll´ ar. Designing network design spaces. In �� , pages 10425–10433, 2020

2020
[72]

Raissi, P

M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physic s-informed neural networks: A deep learning framework for solving forward and inverse problem s involving nonlinear partial diﬀer- ential equations. �� , 378:686–707, 2019

2019
[73]

Fundamentals of recurrent neural ne twork (rnn) and long short-term memory (lstm) network

Alex Sherstinsky. Fundamentals of recurrent neural ne twork (rnn) and long short-term memory (lstm) network. �� , 404:132306, 2020

2020
[74]

Snyder, T

C. Snyder, T. Bengtsson, P. Bickel, and J. Anderson. Obs tacles to high-dimensional particle ﬁltering. �� , 136:4629–4640, 2008

2008
[75]

Performance bounds for particle ﬁlters using the optimal proposal

Chris Snyder, Thomas Bengtsson, and Mathias Morzfeld. Performance bounds for particle ﬁlters using the optimal proposal. �� , 143(11):4750–4761, 2015

2015
[76]

Deep unsuper- vised learning using nonequilibrium thermodynamics

Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranat han, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. In �� , pages 2256–2265. pmlr, 2015

2015
[77]

Song and S

Y. Song and S. Ermon. Generative modeling by estimating gradients of the data distribution. In �� , pages 11918–11930, Vancouver, Canada, Dec. 2019

2019
[78]

Y. Song, J. Sohl-Dickstein, D. P Kingma, A. Kumar, S. Erm on, and B. Poole. Score-based generative modeling through stochastic diﬀerential equati ons. In �� , 2021. 26

2021
[79]

Tan and Q

M. Tan and Q. Le. Eﬃcientnet: Rethinking model scaling f or convolutional neural networks. In �� , pages 6105–6114. PMLR, 2019

2019
[80]

Tsubaki, K

M. Tsubaki, K. Tomii, and J. Sese. Compound–protein int eraction prediction with end-to-end learning of neural networks for graphs and sequences. ��, 35(2):309–318, 2019

2019
[81]

A consistent interpretation of t he stochastic version of the ensemble kalman ﬁlter

Peter Jan van Leeuwen. A consistent interpretation of t he stochastic version of the ensemble kalman ﬁlter. �� , 146(731):2815–2825, 2020

2020
[82]

Wang and J

M. Wang and J. Li. Interpretable predictions of chaotic dynamical systems using dynamical system deep learning. �� , 14(1):3143, 2 2024

2024
[83]

L. Ye, Z. Liu, and Y. Wang. Dual convolutional lstm netwo rk for referring image segmentation. �� , 22(12):3224–3235, 2020

2020
[84]

J. Yin, S. Liu, S. Liang, F. Bao, H. G. Chipilski, D. Lu, an d G. Zhang. A Scalable Real-Time Data Assimilation Framework for Predicting Turbulent Atmo sphere Dynamics. �� ...

2024
[85]

A review of recurrent neural networks: Lstm cells and network architectures

Yong Yu, Xiaosheng Si, Changhua Hu, and Jianxun Zhang. A review of recurrent neural networks: Lstm cells and network architectures. �� , 31(7):1235–1270, 2019

2019
[86]

Iensf: Iterative ensemble sc ore ﬁlter for reducing error in posterior score estimation in nonlinear data assimilation

Z Zhang, F Bao, and G Zhang. Iensf: Iterative ensemble sc ore ﬁlter for reducing error in posterior score estimation in nonlinear data assimilation . ��, page arXiv:2510.20159, 2025

2025 arXiv
[87]

Zhongyi, M

J. Zhongyi, M. Zhu, and Lu Lu. Fourier-mionet: Fourier- enhanced multiple-input neural op- erators for multiphase modeling of geological carbon seque stration. �� , 251:110392, 2024

2024
[88]

M. Zhu, S. Feng, Y. Lin, and Lu Lu. Fourier-deeponet: Fou rier-enhanced deep operator networks for full waveform inversion with improved accuracy, genera lizability, and robustness. �� , 416:116300, 2023. 27

2023

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.

[1] [1]

Allen, S

Anna A. Allen, S. Markou, W. Tebbutt, J. Requeima, W. P. Br uinsma, T. R. Andersson, M. Herzog, N. D. Lane, M .Chantry, J. S. Hosking, and R. E. Turn er. End-to-end data-driven weather prediction. ��, 641(8065):1172–1179, 5 2025. 21

2025

[2] [2]

Andrieu, A

C. Andrieu, A. Doucet, and R. Holenstein. Particle marko v chain monte carlo methods. �� , 72(3):269–342, 2010

2010

[3] [3]

Archibald and F

R. Archibald and F. Bao. A kernel learning method for back ward sde ﬁlter. �� , 2022

2022

[4] [4]

Archibald, F

R. Archibald, F. Bao, and X. Tu. A direct method for parame ter estimations. �� , 398:108871, 2019

2019

[5] [5]

F. Bao, Y. Cao, and X. Han. Forward backward doubly stocha stic diﬀerential equations and optimal ﬁltering of diﬀusion processes. �� , 18(3):635– 661, 2020

2020

[6] [6]

F. Bao, Y. Cao, and J. Yong. Data informed solution estima tion for forward backward stochastic diﬀerential equations. �� , 19:439–464, 2021

2021

[7] [7]

F. Bao, Y. Cao, and W. Zhao. A backward doubly stochastic d iﬀerential equation approach for nonlinear ﬁltering problems. �� , 23:1573–1601, 201

[8] [8]

F. Bao, H. G. Chipilski, S. Liang, G. Zhang, and J. S. Whita ker. Nonlinear ensemble ﬁltering with diﬀusion models: Application to the surface quasigeost rophic dynamics. �� , 153(7):1155–1169, 2025

2025