pith. machine review for the scientific record. sign in

arxiv: 2604.27379 · v1 · submitted 2026-04-30 · 💻 cs.CL · cs.LG

Recognition: unknown

Proactive Dialogue Model with Intent Prediction

Yang Luo
Authors on Pith no claims yet

Pith reviewed 2026-05-07 08:24 UTC · model grok-4.3

classification 💻 cs.CL cs.LG
keywords proactive dialogueintent predictionbayesian networkdialogue systemsmultiwozprompt injectionintent coveragetask-oriented dialogue
0
0 comments X

The pith

A lightweight Bayesian model of intent order, added to the system prompt, makes dialogue models anticipate user goals and cover them in fewer turns.

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

The paper establishes that dialogue models can anticipate upcoming user intents by learning transition patterns from existing data and receiving that information as part of their prompt. A Temporal Bayesian Network trained on MultiWOZ 2.2 supplies the prior probabilities for which intents are likely to appear next. When this prior is included in the system prompt, the model produces responses that cover user goals more thoroughly and in fewer turns. The improvement is measured in replay experiments where the guided model reaches 75 percent coverage after 2.73 turns on average compared with 3.95 turns for the baseline. This matters for practical systems because it adds proactivity to any existing language model at essentially zero extra training cost.

Core claim

The core discovery is that a lightweight intent-transition prior, instantiated as a Temporal Bayesian Network trained on per-turn intent annotations from MultiWOZ 2.2, can be injected into the system prompt to steer generation toward more proactive and efficient dialogue behavior. This yields a Recall@5 of 0.787 for next-intent prediction on held-out turns and, in ground-truth replays of 200 dialogues, raises Coverage AUC from 0.742 to 0.856 while cutting the turns to 75% coverage from 3.95 to 2.73. The method requires no modification to the underlying language model.

What carries the argument

A Temporal Bayesian Network (T-BN) that encodes intent transition probabilities and is supplied to the language model inside the system prompt at inference time.

If this is right

  • BN-guided generation improves Coverage AUC from 0.742 to 0.856.
  • The number of turns required to reach 75% intent coverage drops from 3.95 to 2.73.
  • The T-BN predicts next intents with Recall@5 = 0.787 and MRR = 0.576 on held-out user turns.
  • Proactive and efficient behavior is obtained without modifying or retraining the base language model.

Where Pith is reading between the lines

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

  • Similar priors could be derived from other dialogue datasets to handle different task domains or languages.
  • The approach opens the possibility of dynamically updating the prior from ongoing conversations to personalize the model over time.
  • If the gains hold in open-ended settings, task-oriented dialogue systems could become noticeably less frustrating for users by finishing goals faster.

Load-bearing premise

Intent transition patterns learned from the MultiWOZ 2.2 corpus will match the order in which actual users reveal their intents, and that inserting these probabilities into the prompt will reliably produce more proactive replies without creating new inconsistencies or errors.

What would settle it

A live deployment where the BN-guided model interacts with real users and the number of turns and coverage metrics are compared against an unguided version under the same conditions.

Figures

Figures reproduced from arXiv: 2604.27379 by Yang Luo.

Figure 1
Figure 1. Figure 1: Overview of the proposed pipeline. Stage 1 abstracts a corpus of MultiWOZ dialogues into a binary turn–intent matrix M. Stage 2 lifts M into consecutive USER-turn pairs and fits a Temporal Bayesian Network B with NOTEARS under a forward-only tabu-edge constraint. At runtime, each user utterance u is grounded in B via top-k embedding similarity; the posterior P(xt+1 | O(u)) is thresholded, formatted as a ca… view at source ↗
read the original abstract

Dialogue models are inherently reactive, responding to the current user turn without anticipating upcoming intents, which leads to redundant interactions in multi-intent settings. We address this limitation by introducing a lightweight intent-transition prior derived from dialogue data and injected into the system prompt at inference time. We instantiate this prior using a Temporal Bayesian Network (T-BN) trained on per-turn intent annotations in MultiWOZ 2.2. The T-BN achieves Recall@5 = 0.787 and MRR = 0.576 on 1,071 held-out USER-turn pairs. In a ground-truth replay over 200 dialogues, BN-guided generation improves Coverage AUC from 0.742 to 0.856 and reduces the number of turns required to reach 75% intent coverage from 3.95 to 2.73. These results show that lightweight intent-transition guidance enables more proactive and efficient dialogue behavior without modifying the underlying language model.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The paper claims that dialogue models can be made more proactive by deriving a lightweight intent-transition prior as a Temporal Bayesian Network (T-BN) from per-turn intent annotations in MultiWOZ 2.2 and injecting it into the system prompt at inference time. The T-BN reports Recall@5 = 0.787 and MRR = 0.576 on 1,071 held-out user-turn pairs. In a ground-truth replay over 200 dialogues, BN-guided generation raises Coverage AUC from 0.742 to 0.856 and lowers the turns needed to reach 75% intent coverage from 3.95 to 2.73, enabling more proactive behavior without modifying the underlying language model.

Significance. If the central results hold, the work demonstrates a training-free, low-overhead technique for steering LLMs toward proactive multi-intent dialogue using an existing-data prior. This could meaningfully reduce redundant turns in task-oriented settings. The minimalism of the method (no fine-tuning, prompt-only injection) is a clear strength and supports reproducibility if implementation details are supplied.

major comments (2)
  1. [Abstract] Abstract (results paragraph): The headline gains in Coverage AUC and turns-to-75%-coverage are measured solely by the appearance of annotated intents in generated turns. No automatic or human evaluation of response coherence, relevance, or hallucination is reported, which is load-bearing for the claim of improved proactive behavior; prompt injection of a noisy prior (Recall@5 = 0.787) could degrade surface quality without the coverage metrics detecting it.
  2. [Experimental setup] Experimental setup (held-out evaluation and replay): The manuscript provides no description of the train/held-out split for the 1,071 user-turn pairs, the exact language model and decoding parameters used in the 200-dialogue replay, or the precise prompt template for T-BN injection. These omissions prevent verification that the reported deltas are attributable to the prior rather than implementation artifacts.
minor comments (3)
  1. [Abstract] The abstract states the T-BN is trained on MultiWOZ 2.2 but does not specify whether intent labels were used as-is or post-processed, nor the exact structure of the temporal dependencies in the Bayesian network.
  2. [Results] No statistical significance tests or variance estimates accompany the Coverage AUC or turns-to-coverage numbers, making it impossible to judge whether the observed deltas (0.114 AUC, 1.22 turns) are reliable.
  3. [Evaluation] The paper should clarify whether the replay uses ground-truth user utterances or model-generated ones, as this choice directly affects how much the coverage metric reflects proactive steering versus simple replay fidelity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important aspects of evaluation and reproducibility that we have addressed in the revised manuscript. Below we respond point by point to the major comments.

read point-by-point responses

  1. Referee: [Abstract] Abstract (results paragraph): The headline gains in Coverage AUC and turns-to-75%-coverage are measured solely by the appearance of annotated intents in generated turns. No automatic or human evaluation of response coherence, relevance, or hallucination is reported, which is load-bearing for the claim of improved proactive behavior; prompt injection of a noisy prior (Recall@5 = 0.787) could degrade surface quality without the coverage metrics detecting it.

    Authors: We agree that intent coverage alone does not fully establish that the generated responses remain coherent and relevant. Our ground-truth replay isolates the effect of the T-BN prior on system behavior while holding user turns fixed, and the coverage metrics directly quantify the reduction in redundant turns. Nevertheless, we acknowledge that additional quality checks are necessary to rule out degradation from the noisy prior. In the revised manuscript we add (i) automatic metrics (ROUGE-L and BERTScore against reference system responses on the same replay set) and (ii) a human evaluation on 50 sampled dialogues rating coherence, relevance, and absence of hallucination on a 5-point Likert scale. These results will be reported in a new subsection of the Experiments section. revision: yes

  2. Referee: [Experimental setup] Experimental setup (held-out evaluation and replay): The manuscript provides no description of the train/held-out split for the 1,071 user-turn pairs, the exact language model and decoding parameters used in the 200-dialogue replay, or the precise prompt template for T-BN injection. These omissions prevent verification that the reported deltas are attributable to the prior rather than implementation artifacts.

    Authors: We thank the referee for identifying these reproducibility gaps. The 1,071 held-out user-turn pairs were obtained by randomly selecting 50 complete dialogues from the MultiWOZ 2.2 test set (ensuring no dialogue overlap with the data used to construct the T-BN). The replay experiments used GPT-3.5-turbo with temperature = 0.7, top_p = 0.95, and max_tokens = 128. The exact prompt template injected the T-BN probabilities as: “You are a proactive task-oriented assistant. The following intent-transition prior is available: [list of top-5 next intents with probabilities]. Anticipate the user’s next intent and respond accordingly.” A new “Experimental Setup” subsection has been added to the revised manuscript containing the full split procedure, model name, decoding hyperparameters, and verbatim prompt template. revision: yes

Circularity Check

0 steps flagged

No circularity: T-BN prior trained externally, coverage gains measured independently in replay

full rationale

The T-BN is trained on MultiWOZ 2.2 per-turn intent annotations and its predictive accuracy is reported on a separate held-out set of 1,071 user-turn pairs (Recall@5 = 0.787). The headline results (Coverage AUC 0.742→0.856, turns-to-75% 3.95→2.73) come from a distinct ground-truth replay experiment over 200 dialogues that measures how well LM-generated responses align with the ground-truth intent sequence when the prior is injected into the prompt. These downstream metrics are not algebraically or statistically forced by the BN parameters themselves; they depend on the external LM's response to the prompt and on the replay metric definition. No self-citations, self-definitional equations, fitted-input-renamed-as-prediction, or ansatz smuggling appear in the derivation chain. The evaluation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on fitting transition probabilities from an existing dialogue corpus and assuming those probabilities remain useful when injected into an unrelated language model.

free parameters (1)
  • Intent transition probabilities
    Estimated directly from per-turn intent annotations in MultiWOZ 2.2; these probabilities constitute the learned prior.
axioms (1)
  • domain assumption User intents exhibit stable temporal dependencies that can be captured by a first-order Bayesian network
    The T-BN construction presupposes that intent sequences are sufficiently Markovian for the network to generalize.

pith-pipeline@v0.9.0 · 5444 in / 1362 out tokens · 84106 ms · 2026-05-07T08:24:38.827829+00:00 · methodology

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

38 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Dialogue language model with large-scale persona data engineering

    Mengze Hong, Chen Jason Zhang, Chaotao Chen, Rongzhong Lian, and Di Jiang. Dialogue language model with large-scale persona data engineering. In Weizhu Chen, Yi Yang, Moham- mad Kachuee, and Xue-Yong Fu, editors,Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Techno...

    2025

  2. [2]

    Association for Computational Linguistics

  3. [3]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwi...

    1901

  4. [4]

    QualBench: Benchmarking Chinese LLMs with localized professional qualifications for vertical domain evaluation

    Mengze Hong, Wailing Ng, Chen Jason Zhang, and Di Jiang. QualBench: Benchmarking Chinese LLMs with localized professional qualifications for vertical domain evaluation. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors, Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 59...

    2025

  5. [5]

    Orchestration-free customer service automation: A privacy-preserving and flowchart-guided framework

    Mengze Hong, Chen Jason Zhang, Zichang Guo, Hanlin Gu, Di Jiang, and Qing Li. Orchestration-free customer service automation: A privacy-preserving and flowchart-guided framework. InProceedings of the ACM Web Conference 2026, WWW ’26, page 8138–8149, New York, NY , USA, 2026. Association for Computing Machinery

    2026

  6. [6]

    How learners engage with an llm-based pedagogical conversational agent during music form analysis

    Lingxi Jin, Kyuwon Kim, Baicheng Lin, Mengze Hong, and Hyo-Jeong So. How learners engage with an llm-based pedagogical conversational agent during music form analysis. In Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems, CHI EA ’26, New York, NY , USA, 2026. Association for Computing Machinery

    2026

  7. [7]

    Exploring the impact of an llm-powered teachable agent on learning gains and cognitive load in music education, 2025

    Lingxi Jin, Baicheng Lin, Mengze Hong, Kun Zhang, and Hyo-Jeong So. Exploring the impact of an llm-powered teachable agent on learning gains and cognitive load in music education, 2025

    2025

  8. [8]

    Predicting the next app that you are going to use

    Ricardo Baeza-Yates, Di Jiang, Fabrizio Silvestri, and Beverly Harrison. Predicting the next app that you are going to use. InProceedings of the Eighth ACM International Conference on Web Search and Data Mining, WSDM ’15, page 285–294, New York, NY , USA, 2015. Association for Computing Machinery

    2015

  9. [9]

    Panorama: A semantic-aware application search framework

    Di Jiang, Jan V osecky, Kenneth Wai-Ting Leung, and Wilfred Ng. Panorama: A semantic-aware application search framework. InProceedings of the 16th international conference on extending database technology, pages 371–382, 2013

    2013

  10. [10]

    Discovering new intents via constrained deep adaptive clustering with cluster refinement

    Ting-En Lin, Hua Xu, and Hanlei Zhang. Discovering new intents via constrained deep adaptive clustering with cluster refinement. InAAAI, 2020

    2020

  11. [11]

    Neural-bayesian program learning for few-shot dialogue intent parsing, 2024

    Mengze Hong, Di Jiang, Yuanfeng Song, and Chen Jason Zhang. Neural-bayesian program learning for few-shot dialogue intent parsing, 2024

    2024

  12. [12]

    Llm-in-the-loop: Replicating human insight with llms for better machine learning applications

    Mengze Hong, Wailing Ng, Chen Jason Zhang, Yifei Wang, Yuanfeng Song, and Di Jiang. Llm-in-the-loop: Replicating human insight with llms for better machine learning applications. TechRxiv, 2025(0528), 2025

    2025

  13. [13]

    Dial-in LLM: Human-aligned LLM-in-the-loop intent clustering for customer service dialogues

    Mengze Hong, Wailing Ng, Chen Jason Zhang, Yuanfeng Song, and Di Jiang. Dial-in LLM: Human-aligned LLM-in-the-loop intent clustering for customer service dialogues. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors,Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 5885– 5...

    2025

  14. [14]

    Efficient intent detection with dual sentence encoders

    Iñigo Casanueva, Tadas Temˇcinas, Daniela Gerz, Matthew Henderson, and Ivan Vuli´c. Efficient intent detection with dual sentence encoders. InProceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, pages 38–45, 2020

    2020

  15. [15]

    A survey of joint intent detection and slot filling models in natural language understanding.ACM Computing Surveys, 55(8):1–38, 2022

    Henry Weld, Xiaoqi Huang, Siqu Long, Josiah Poon, and Soyeon Caren Han. A survey of joint intent detection and slot filling models in natural language understanding.ACM Computing Surveys, 55(8):1–38, 2022

    2022

  16. [16]

    Multiwoz 2.2 : A dialogue dataset with additional annotation corrections and state tracking baselines

    Xiaoxue Zang, Abhinav Rastogi, Srinivas Sunkara, Raghav Gupta, Jianguo Zhang, and Jindong Chen. Multiwoz 2.2 : A dialogue dataset with additional annotation corrections and state tracking baselines. pages 109–117, 2020. 7

    2020

  17. [17]

    Transferable multi-domain state generator for task-oriented dialogue systems

    Chien-Sheng Wu, Andrea Madotto, Ehsan Hosseini-Asl, Caiming Xiong, Richard Socher, and Pascale Fung. Transferable multi-domain state generator for task-oriented dialogue systems. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 808–819, 2019

    2019

  18. [18]

    Trippy: A triple copy strategy for value independent neural dialog state tracking

    Michael Heck, Carel van Niekerk, Nurul Lubis, Christian Geishauser, Hsien-Chin Lin, Marco Moresi, and Milica Gaši´c. Trippy: A triple copy strategy for value independent neural dialog state tracking. pages 35–44, 2020

    2020

  19. [19]

    A simple language model for task-oriented dialogue

    Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu, Semih Yavuz, and Richard Socher. A simple language model for task-oriented dialogue. InProceedings of the 34th International Conference on Neural Information Processing Systems, page 13, Red Hook, NY , USA, 2020. Curran Associates Inc

    2020

  20. [20]

    Soloist: Building task bots at scale with transfer learning and machine teaching.Transactions of the Association for Computational Linguistics, 9:807–824, 2021

    Baolin Peng, Chunyuan Li, Jinchao Li, Shahin Shayandeh, Lars Liden, and Jianfeng Gao. Soloist: Building task bots at scale with transfer learning and machine teaching.Transactions of the Association for Computational Linguistics, 9:807–824, 2021

    2021

  21. [21]

    Multi-task pre-training for plug-and-play task-oriented dialogue system

    Yixuan Su, Lei Shu, Elman Mansimov, Arshit Gupta, Deng Cai, Yi-An Lai, and Yi Zhang. Multi-task pre-training for plug-and-play task-oriented dialogue system. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4661–4676, Dublin, Ireland, 2022. Association for Computational Linguistics

    2022

  22. [22]

    Federated heterogeneous language model optimization for hybrid automatic speech recognition

    Mengze Hong, Yi Gu, Di Jiang, Hanlin Gu, Chen Jason Zhang, Lu Wang, and Zhiyang Su. Federated heterogeneous language model optimization for hybrid automatic speech recognition. InICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 18477–18481, 2026

    2026

  23. [23]

    ASR-EC bench- mark: Evaluating large language models on Chinese ASR error correction

    Victor Junqiu Wei, Weicheng Wang, Di Jiang, Yuanfeng Song, and Lu Wang. ASR-EC bench- mark: Evaluating large language models on Chinese ASR error correction. In Saloni Potdar, Lina Rojas-Barahona, and Sebastien Montella, editors,Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 1567–1575, Suzhou ...

    2025

  24. [24]

    InfantCryNet: A data-driven framework for intelligent analysis of infant cries

    Mengze Hong, Chen Jason Zhang, Lingxiao Yang, Yuanfeng SONG, and Di Jiang. InfantCryNet: A data-driven framework for intelligent analysis of infant cries. In Vu Nguyen and Hsuan-Tien Lin, editors,Proceedings of the 16th Asian Conference on Machine Learning, volume 260 of Proceedings of Machine Learning Research, pages 845–857. PMLR, 05–08 Dec 2025

    2025

  25. [25]

    Contextu- alized token discrimination for speech search query correction, 2025

    Junyu Lu, Di Jiang, Mengze Hong, Victor Junqiu Wei, Qintian Guo, and Zhiyang Su. Contextu- alized token discrimination for speech search query correction, 2025

    2025

  26. [26]

    Xun Zheng, Bryon Aragam, Pradeep Ravikumar, and Eric P. Xing. Dags with no tears: continuous optimization for structure learning. InProceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS’18, page 9492–9503, Red Hook, NY , USA, 2018. Curran Associates Inc

    2018

  27. [27]

    Xun Zheng, Chen Dan, Bryon Aragam, Pradeep Ravikumar, and Eric P. Xing. Learning sparse nonparametric dags. InProceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics(AISTATS), pages 3414–3425, 2020

    2020

  28. [28]

    Dag-gnn: Dag structure learning with graph neural networks.ArXiv, abs/1904.10098, 2019

    Yue Yu, Jie Chen, Tian Gao, and Mo Yu. Dag-gnn: Dag structure learning with graph neural networks.ArXiv, abs/1904.10098, 2019

    work page arXiv 1904

  29. [29]

    Gradient- based neural dag learning, 2020

    Sébastien Lachapelle, Philippe Brouillard, Tristan Deleu, and Simon Lacoste-Julien. Gradient- based neural dag learning, 2020

    2020

  30. [30]

    Learning the structure of dynamic probabilistic networks

    Nir Friedman, Kevin Murphy, and Stuart Russell. Learning the structure of dynamic probabilistic networks. InProceedings of the Fourteenth Conference on Uncertainty in Artificial Intelligence, page 139–147, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers Inc

    1998

  31. [31]

    Murphy.Dynamic Bayesian Networks: Representation, Inference and Learning

    Kevin P. Murphy.Dynamic Bayesian Networks: Representation, Inference and Learning. PhD thesis, University of California, Berkeley, 2002. 8

    2002

  32. [32]

    Dynotears: Structure learning from time-series data

    Roxana Pamfil, Nisara Sriwattanaworachai, Shaan Desai, Philip Pilgerstorfer, Konstantinos Georgatzis, Paul Beaumont, and Bryon Aragam. Dynotears: Structure learning from time-series data. InAISTATS, 2020

    2020

  33. [33]

    Augmenting compliance- guaranteed customer service chatbots: Context-aware knowledge expansion with large language models

    Mengze Hong, Chen Jason Zhang, Di Jiang, and Yuanqin He. Augmenting compliance- guaranteed customer service chatbots: Context-aware knowledge expansion with large language models. In Saloni Potdar, Lina Rojas-Barahona, and Sebastien Montella, editors,Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pa...

    2025

  34. [34]

    Efficient Guided Generation for Large Language Models

    Brandon T. Willard and Rémi Louf. Efficient guided generation for large language models. ArXiv, abs/2307.09702, 2023

    work page internal anchor Pith review arXiv 2023

  35. [35]

    Yuwei Zhang, Haode Zhang, Li-Ming Zhan, Xiao-Ming Wu, and Albert Y .S. Lam. New intent discovery with pre-training and contrastive learning. InACL, 2022

    2022

  36. [36]

    Large language models enable few-shot clustering

    Vijay Viswanathan, Kiril Gashteovski, Carolin Lawrence, Tongshuang Wu, and Graham Neubig. Large language models enable few-shot clustering. InTACL, 2024

    2024

  37. [37]

    Proactive conversational agents

    Lizi Liao, Ryuichi Takanobu, Yunshan Ma, Xun Yang, Minlie Huang, and Tat-Seng Chua. Proactive conversational agents. InWSDM Tutorial, 2023

    2023

  38. [38]

    A survey on proactive dialogue systems: Problems, methods, and prospects.arXiv preprint arXiv:2305.02750, 2023

    Yang Deng, Lizi Liao, Liang Chen, Hongru Wang, Wenqiang Lei, and Tat-Seng Chua. A survey on proactive dialogue systems: Problems, methods, and prospects.arXiv preprint arXiv:2305.02750, 2023. 9

    work page arXiv 2023