OrthoPhys: Physically Plausible Video Generation with Orthogonal-View Geometry Guidance

Cong Wang; Hanxin Zhu; Jiayi Luo; Long Chen; Xiao Tang; Xin Jin; Zhibo Chen

OrthoPhys makes generated video motion more physically plausible by first locking four synchronized orthogonal views of the foreground.

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 · grok-4.5

2026-07-13 22:29 UTC pith:2RKYQAIG

load-bearing objection We only have OrthoPhys’s abstract; the supplied full text is a different NLP paper, so the physical-plausibility claims cannot be checked. the 3 major comments →

arxiv 2603.18639 v3 pith:2RKYQAIG submitted 2026-03-19 cs.CV

OrthoPhys: Physically Plausible Video Generation with Orthogonal-View Geometry Guidance

Cong Wang , Hanxin Zhu , Xiao Tang , Jiayi Luo , Xin Jin , Long Chen , Zhibo Chen This is my paper

classification cs.CV

keywords video generationphysical plausibilityorthogonal viewsgeometry-enhanced attentionmulti-view guidancePhysMVspatial-temporal coherenceforeground dynamics

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.

Single-camera video generators often look sharp yet produce motion that breaks physical common sense, because each frame is only a 2D projection of dynamics that actually unfold in 3D. OrthoPhys attacks that gap with a two-stage pipeline. The first stage generates four synchronized orthogonal videos of the moving foreground and links them with a geometry-enhanced attention mechanism so the motion stays 3D-coherent and is implicitly tied to physical attributes. Those multi-view foregrounds then serve as rigid guidance while a second stage synthesizes the complete scene and background. To train the system the authors build PhysMV, a 40K-scene dataset with four orthogonal viewpoints per scene (160K sequences total), and report clear gains in physical realism and spatial-temporal coherence over existing generators.

Core claim

Generating synchronized four-view orthogonal videos of foreground dynamics, joined by geometry-enhanced attention, enforces 3D spatial coherence and implicitly grounds motion in physical attributes; using those multi-view foregrounds as rigid guidance then yields complete videos with substantially better physical realism and spatial-temporal coherence than direct unstructured 2D generation.

What carries the argument

Orthogonal-view geometry guidance: a two-stage process that first produces four synchronized orthogonal foreground videos linked by geometry-enhanced cross-view attention, then treats those sequences as rigid constraints when synthesizing the final full video.

Load-bearing premise

The work assumes that forcing four orthogonal 2D views of the same motion to agree is enough to ground that motion in real physical attributes, without any explicit physics rules or dynamics model.

What would settle it

Run OrthoPhys on simple scenes with known physics (free fall, collisions, rigid rotation) and measure whether multi-view-consistent outputs still violate conservation laws, contact, or trajectories at rates similar to single-view baselines; if they do, geometric multi-view agreement is not delivering physical plausibility.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

We only have OrthoPhys’s abstract; the supplied full text is a different NLP paper, so the physical-plausibility claims cannot be checked.

read the letter

The first thing to know: the cache is broken. Metadata and abstract are OrthoPhys (orthogonal multi-view video generation for physical plausibility). The full manuscript text is an unrelated continual-learning intent paper on CLINC150. So this is effectively abstract-only for the titled work. I am not going to pretend I read experiments that are not here.

From the abstract alone, the pitch is clear and useful if true. Physical consistency is still a real gap in generative video. OrthoPhys’s claimed novelty is a two-stage recipe: (1) generate synchronized four-view orthogonal foreground videos with a geometry-enhanced attention mechanism to enforce 3D spatial coherence, then (2) use those as rigid guidance for full-scene synthesis. They also claim PhysMV—40K scenes, four orthogonal views, 160K sequences. That is a concrete methods-plus-data package, not just a slogan. Multi-view consistency and guidance-based synthesis are established ideas; the specific orthogonal-foreground-then-guided-composite framing and the dataset are what would be new if the paper delivers them.

The soft spot is load-bearing and currently uncheckable. The abstract treats multi-view geometric coherence as enough to “implicitly ground the motion in physical attributes.” That may improve 3D consistency; it is not the same as obeying dynamics, contact, or conservation laws. Without architecture details, baselines, ablations, metrics, or failure cases, I cannot tell whether they measured physical realism or just multi-view agreement. Free design choices (exactly four views, how foreground is defined, how PhysMV is built) also sit outside any derivation we can inspect.

Who this is for: people already working on video generation, 4D synthesis, or multi-view consistency who would care about a practical guidance recipe and a multi-view motion dataset. It is not a field-reorganizing result on the abstract’s face—mid-tier methods advance if the missing experiments hold.

Recommendation: do not spend reading-group time on the mismatched PDF. If the real OrthoPhys manuscript appears with solid baselines and honest physics metrics, it deserves a serious referee. On abstract alone I would not cite it, and I would not treat the physical-plausibility claim as established. Engage only after the correct full paper is in hand.

Referee Report

3 major / 0 minor

Summary. Based solely on the supplied abstract, the manuscript claims that OrthoPhys, a two-stage video generation framework, improves physical realism by first synthesizing synchronized four-view orthogonal foreground videos with a geometry-enhanced attention mechanism (to enforce 3D spatial coherence and implicitly ground motion in physical attributes), then using those foregrounds as rigid guidance for full-video synthesis that learns foreground–background interaction. A supporting dataset PhysMV (40K multi-view scenes / 160K sequences) is introduced, and extensive experiments are said to show gains in physical realism and spatio-temporal coherence over existing methods. The full manuscript text provided in the review package, however, is an unrelated empirical study of catastrophic forgetting mitigation for continual intent classification on CLINC150 (ANN/GRU/Transformer backbones with MIR, LwF, HAT and combinations), not OrthoPhys.

Significance. If the OrthoPhys claims held as stated in the abstract—orthogonal multi-view generation plus geometry-enhanced attention yielding measurably more physically plausible video without an explicit physics engine—the work would be a meaningful contribution to video generation and 3D-aware generative modeling, especially with a 160K-sequence multi-view dataset. That significance cannot be assessed from the materials actually supplied: the body text, figures, tables, method definitions, and experiments belong to a different paper (continual NLP learning). No architecture equations, PhysMV construction details, baselines, metrics, or ablations for OrthoPhys are available to evaluate.

major comments (3)

Manuscript identity mismatch: the title/abstract/paper_id (OrthoPhys, arXiv 2603.18639, cs.CV, physically plausible video generation) do not match the full manuscript text, which is a continual-learning NLP study on CLINC150 (catastrophic forgetting with MIR/LwF/HAT; arXiv 2603.18641). Sections, figures (e.g., utterance-length distribution, AA/AF1/BWT curves), methods, and references all belong to the NLP paper. A technical review of OrthoPhys’s central claims is therefore impossible from the provided package.
Because the OrthoPhys body is absent, load-bearing elements cannot be checked: definition of the geometry-enhanced attention across four orthogonal views; how multi-view 2D consistency is argued to imply physical lawfulness without explicit dynamics; PhysMV construction and camera layout; second-stage ‘rigid guidance’ mechanism; quantitative tables, baselines, ablations, and failure cases. The abstract’s premise that synchronized orthogonal views ‘implicitly ground the motion in physical attributes’ remains untestable.
Even at abstract level, the operationalization of ‘physical plausibility’ as multi-view geometric consistency risks circularity (gains on consistency metrics partly restating the training objective). Without the missing method and evaluation sections this cannot be confirmed or refuted; it is flagged only as a correctness-risk that any resubmission must address with physics-oriented metrics beyond multi-view agreement.

Circularity Check

0 steps flagged

No significant circularity: OrthoPhys body is missing (text is an unrelated CLINC150 CL study); neither the abstract premise nor the supplied empirical manuscript reduces a claimed prediction to its inputs by construction.

full rationale

The cacheable prefix labels the paper as OrthoPhys (orthogonal-view video generation for physical plausibility) but the full manuscript text is an unrelated empirical study of catastrophic forgetting on CLINC150 intent classification (ANN/GRU/Transformer with MIR, LwF, HAT). OrthoPhys therefore has no equations, architecture definitions, PhysMV construction details, or evaluation protocol available to walk. The abstract’s design claim—that synchronized four orthogonal views plus geometry-enhanced attention “implicitly grounds the motion in physical attributes”—is a modeling premise, not a closed-form derivation that equates a fitted quantity to a later “prediction.” Without the body, no self-definitional identity, fitted-input-as-prediction, load-bearing self-citation uniqueness theorem, or renamed known result can be exhibited with quotes. The supplied CLINC150 manuscript is a comparative empirical study: it reports measured AA/AF1/BWT under sequential fine-tuning and CL combinations, with no first-principles prediction chain and only ordinary literature citations to MIR/LwF/HAT. Per the analyzer rules, absence of a reducible derivation yields score 0 and empty steps; mild definitional risk that “physical plausibility” might later be operationalized as multi-view consistency cannot be elevated to circularity without quotable reduction from the actual OrthoPhys text.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 2 invented entities

Review is abstract-only for OrthoPhys; free parameters, training losses, and architectural constants are not specified. Load-bearing premises are the domain assumptions that multi-view orthogonal consistency proxies physical plausibility and that four fixed orthogonal views suffice as geometric scaffolding for full-video synthesis.

free parameters (2)

Number of orthogonal views (fixed to 4) = 4
Abstract fixes four orthogonal viewpoints without a derivation that four is necessary or optimal; this is a design choice the method depends on.
PhysMV dataset construction choices (scene count, camera layout, foreground definition)
40K scenes / 160K sequences and the orthogonal camera layout are author-chosen training conditions that determine what 'physical' multi-view consistency the model can learn; values not derived from first principles.

axioms (3)

domain assumption Real-world object motion unfolds in 3D while video is a partial view-dependent projection, and this mismatch is a primary cause of physically inconsistent generated motion.
Stated as the motivating premise in the abstract; plausible but not proven to be the dominant failure mode of current generators.
ad hoc to paper Synchronized orthogonal multi-view generation with geometry-enhanced attention enforces 3D spatial coherence and implicitly grounds motion in physical attributes without an explicit physics model.
Core design claim of stage 1; treats multi-view consistency as a sufficient proxy for physical plausibility.
ad hoc to paper Physically consistent orthogonal foregrounds can serve as rigid guidance so a second stage learns correct foreground-background interaction.
Stage-2 premise in the abstract; assumes guidance remains rigid and transfers physical consistency into the final single-view video.

invented entities (2)

OrthoPhys two-stage framework (geometry-enhanced attention across four orthogonal views + guided full-video synthesis) no independent evidence
purpose: Enforce physical plausibility and spatio-temporal coherence in generated video via multi-view geometric guidance.
Named method introduced by the paper; no independent external validation available in the provided text.
PhysMV dataset (40K multi-view scenes / 160K sequences) no independent evidence
purpose: Support orthogonal-view training for physically consistent video generation.
New dataset claimed in the abstract; release status and labeling protocol not verifiable from available text.

pith-pipeline@v1.1.0-grok45 · 10754 in / 2881 out tokens · 31117 ms · 2026-07-13T22:29:07.558800+00:00 · methodology

0 comments

read the original abstract

Recent progress in video generation has led to substantial improvements in visual fidelity, yet ensuring physically consistent motion remains a fundamental challenge. Intuitively, this limitation can be attributed to the fact that real-world object motion unfolds in three-dimensional space, while video observations provide only partial, view-dependent projections of such dynamics. To address these issues, we propose OrthoPhys, a two-stage framework that leverages orthogonal-view geometry guidance to enforce physical plausibility. Instead of directly generating unstructured 2D videos, our first stage generates synchronized, four-view orthogonal videos of the foreground dynamics. By incorporating a geometry-enhanced attention mechanism across these orthogonal views, this stage effectively enforces 3D spatial coherence and implicitly grounds the motion in physical attributes. In the second stage, these physically consistent orthogonal foregrounds serve as rigid guidance to synthesize the final complete video, seamlessly learning the interaction between foreground dynamics and the background context. To support this orthogonal-view training paradigm, we construct PhysMV, a dataset containing 40K scenes, each consisting of four orthogonal viewpoints, resulting in a total of 160K video sequences. Extensive experiments demonstrate that OrthoPhys significantly improves physical realism and spatial-temporal coherence over existing video generation methods. Project page: https://anonymous.4open.science/w/Phys4D/.

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 6 canonical work pages · 1 internal anchor

[1]

IEEE transactions on pattern analysis and machine intelligence, 44(7), pp.3366-3385

DeLange,M.,Aljundi,R.,Masana,M.,Parisot,S.,Jia,X.,Leonardis,A.,Slabaugh,G.andTuytelaars,T.,2021.Acontinuallearningsurvey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence, 44(7), pp.3366-3385

2021
[2]

Advances in neural information processing systems, 32

Aljundi,R.,Belilovsky,E.,Tuytelaars,T.,Charlin,L.,Caccia,M.,Lin,M.andPage-Caccia,L.,2019.Onlinecontinuallearningwithmaximal interfered retrieval. Advances in neural information processing systems, 32

2019
[3]

and Ranzato, M.A., 2017

Lopez-Paz, D. and Ranzato, M.A., 2017. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30

2017
[4]

Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A.andHassabis,D.,2017.Overcomingcatastrophicforgettinginneuralnetworks.Proceedingsofthenationalacademyofsciences,114(13), pp.3521-3526

2017
[5]

and Hoiem, D., 2017

Li, Z. and Hoiem, D., 2017. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12), pp.2935- 2947

2017
[6]

and Hadsell, R., 2016

Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R. and Hadsell, R., 2016. Progressive neural networks. arXiv preprint arXiv:1606.04671

Pith/arXiv arXiv 2016
[7]

and Lazebnik, S., 2018

Mallya, A. and Lazebnik, S., 2018. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (pp. 7765-7773)

2018
[8]

and Karatzoglou, A., 2018, July

Serra, J., Suris, D., Miron, M. and Karatzoglou, A., 2018, July. Overcoming catastrophic forgetting with hard attention to the task. In International conference on machine learning (pp. 4548-4557). PMLR

2018
[9]

I., Kemker, R., Part, J

Parisi, G. I., Kemker, R., Part, J. L., Kanan, C., & Wermter, S. (2019). Continual lifelong learning with neural networks: A review. Neural Networks, 113, 54-71. https://doi.org/10.1016/j.neunet.2019.01.012

doi:10.1016/j.neunet.2019.01.012 2019
[10]

https://doi.org/10.48550/arxiv.2401.16386

Zhou,D.,Sun,H.,Ning,J.,Ye,H.,&Zhan,D.(2024b).ContinualLearningwithPre-TrainedModels:ASurvey.arXiv(CornellUniversity). https://doi.org/10.48550/arxiv.2401.16386

Pith/arXiv arXiv doi:10.48550/arxiv.2401.16386
[11]

Biesialska, M., Biesialska, K., & Costa-Jussà, M. R. (2020). Continual Lifelong Learning in Natural Language Processing: A Survey. ACL Anthology. https://doi.org/10.18653/v1/2020.coling-main.574

doi:10.18653/v1/2020.coling-main.574 2020
[12]

Ke, Z., & Liu, B. (2022). Continual Learning of Natural Language Processing Tasks: a survey. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2211.12701

Pith/arXiv arXiv doi:10.48550/arxiv.2211.12701 2022
[13]

De Masson D’Autume, C., Ruder, S., Kong, L., & Yogatama, D. (2019). Episodic memory in lifelong language learning. arXiv (Cornell University). https://doi.org/10.48550/arxiv.1906.01076

Pith/arXiv arXiv doi:10.48550/arxiv.1906.01076 2019
[14]

Sun, F., Ho, C., & Lee, H. (2019b). LAMOL: LAnguage MOdeling for Lifelong Language Learning. arXiv (Cornell University). https://doi.org/10.48550/arxiv.1909.03329

Pith/arXiv arXiv doi:10.48550/arxiv.1909.03329 1909
[15]

ACL Anthology

Huang,Y.,Zhang,Y.,Chen,J.,Wang,X.,&Yang,D.(2021b).ContinualLearningforTextClassificationwithInformationDisentanglement Based Regularization. ACL Anthology. https://doi.org/10.18653/v1/2021.naacl-main.218

work page doi:10.18653/v1/2021.naacl-main.218 2021
[16]

Paul, D., Sorokin, D., & Gaspers, J. (2022). Class Incremental Learning for Intent Classification with Limited or No Old Data. ACL Anthology, 16-25. https://doi.org/10.18653/v1/2022.evonlp-1.4

work page doi:10.18653/v1/2022.evonlp-1.4 2022
[17]

Luoyiching, C., Li, Y., Li, Y., Li, R., Zheng, H., Zhou, N., & Su, H. (2023). Prompt learning with knowledge memorizing prototypes for generalized Few-Shot intent detection. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2309.04971

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2309.04971 2023
[18]

Zhang, X., Jiang, M., Chen, H., Zheng, J., & Pan, Z. (2022). Incorporating geometry knowledge into an incremental learning structure for few-shot intent recognition. Knowledge-Based Systems, 251, 109296. https://doi.org/10.1016/j.knosys.2022.109296

work page doi:10.1016/j.knosys.2022.109296 2022
[19]

ACL Anthology

Song,X.,Mou,Y.,He,K.,Qiu,Y.,Zhao,J.,Wang,P.,&Xu,W.(2023).ContinualGeneralizedIntentDiscovery:MarchingTowardsDynamic and Open-world Intent Recognition. ACL Anthology. https://doi.org/10.18653/v1/2023.findings-emnlp.289

work page doi:10.18653/v1/2023.findings-emnlp.289 2023
[20]

CLINC150. 2020. UCI Machine Learning Repository. https://doi.org/10.24432/C5MP58. First Author et al.:Preprint submitted to ElsevierPage 19 of 19

work page doi:10.24432/c5mp58 2020

[1] [1]

IEEE transactions on pattern analysis and machine intelligence, 44(7), pp.3366-3385

DeLange,M.,Aljundi,R.,Masana,M.,Parisot,S.,Jia,X.,Leonardis,A.,Slabaugh,G.andTuytelaars,T.,2021.Acontinuallearningsurvey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence, 44(7), pp.3366-3385

2021

[2] [2]

Advances in neural information processing systems, 32

Aljundi,R.,Belilovsky,E.,Tuytelaars,T.,Charlin,L.,Caccia,M.,Lin,M.andPage-Caccia,L.,2019.Onlinecontinuallearningwithmaximal interfered retrieval. Advances in neural information processing systems, 32

2019

[3] [3]

and Ranzato, M.A., 2017

Lopez-Paz, D. and Ranzato, M.A., 2017. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30

2017

[4] [4]

Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A.andHassabis,D.,2017.Overcomingcatastrophicforgettinginneuralnetworks.Proceedingsofthenationalacademyofsciences,114(13), pp.3521-3526

2017

[5] [5]

and Hoiem, D., 2017

Li, Z. and Hoiem, D., 2017. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12), pp.2935- 2947

2017

[6] [6]

and Hadsell, R., 2016

Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R. and Hadsell, R., 2016. Progressive neural networks. arXiv preprint arXiv:1606.04671

Pith/arXiv arXiv 2016

[7] [7]

and Lazebnik, S., 2018

Mallya, A. and Lazebnik, S., 2018. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (pp. 7765-7773)

2018

[8] [8]

and Karatzoglou, A., 2018, July

Serra, J., Suris, D., Miron, M. and Karatzoglou, A., 2018, July. Overcoming catastrophic forgetting with hard attention to the task. In International conference on machine learning (pp. 4548-4557). PMLR

2018

[9] [9]

I., Kemker, R., Part, J

Parisi, G. I., Kemker, R., Part, J. L., Kanan, C., & Wermter, S. (2019). Continual lifelong learning with neural networks: A review. Neural Networks, 113, 54-71. https://doi.org/10.1016/j.neunet.2019.01.012

doi:10.1016/j.neunet.2019.01.012 2019

[10] [10]

https://doi.org/10.48550/arxiv.2401.16386

Zhou,D.,Sun,H.,Ning,J.,Ye,H.,&Zhan,D.(2024b).ContinualLearningwithPre-TrainedModels:ASurvey.arXiv(CornellUniversity). https://doi.org/10.48550/arxiv.2401.16386

Pith/arXiv arXiv doi:10.48550/arxiv.2401.16386

[11] [11]

Biesialska, M., Biesialska, K., & Costa-Jussà, M. R. (2020). Continual Lifelong Learning in Natural Language Processing: A Survey. ACL Anthology. https://doi.org/10.18653/v1/2020.coling-main.574

doi:10.18653/v1/2020.coling-main.574 2020

[12] [12]

Ke, Z., & Liu, B. (2022). Continual Learning of Natural Language Processing Tasks: a survey. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2211.12701

Pith/arXiv arXiv doi:10.48550/arxiv.2211.12701 2022

[13] [13]

De Masson D’Autume, C., Ruder, S., Kong, L., & Yogatama, D. (2019). Episodic memory in lifelong language learning. arXiv (Cornell University). https://doi.org/10.48550/arxiv.1906.01076

Pith/arXiv arXiv doi:10.48550/arxiv.1906.01076 2019

[14] [14]

Sun, F., Ho, C., & Lee, H. (2019b). LAMOL: LAnguage MOdeling for Lifelong Language Learning. arXiv (Cornell University). https://doi.org/10.48550/arxiv.1909.03329

Pith/arXiv arXiv doi:10.48550/arxiv.1909.03329 1909

[15] [15]

ACL Anthology

Huang,Y.,Zhang,Y.,Chen,J.,Wang,X.,&Yang,D.(2021b).ContinualLearningforTextClassificationwithInformationDisentanglement Based Regularization. ACL Anthology. https://doi.org/10.18653/v1/2021.naacl-main.218

work page doi:10.18653/v1/2021.naacl-main.218 2021

[16] [16]

Paul, D., Sorokin, D., & Gaspers, J. (2022). Class Incremental Learning for Intent Classification with Limited or No Old Data. ACL Anthology, 16-25. https://doi.org/10.18653/v1/2022.evonlp-1.4

work page doi:10.18653/v1/2022.evonlp-1.4 2022

[17] [17]

Luoyiching, C., Li, Y., Li, Y., Li, R., Zheng, H., Zhou, N., & Su, H. (2023). Prompt learning with knowledge memorizing prototypes for generalized Few-Shot intent detection. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2309.04971

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2309.04971 2023

[18] [18]

Zhang, X., Jiang, M., Chen, H., Zheng, J., & Pan, Z. (2022). Incorporating geometry knowledge into an incremental learning structure for few-shot intent recognition. Knowledge-Based Systems, 251, 109296. https://doi.org/10.1016/j.knosys.2022.109296

work page doi:10.1016/j.knosys.2022.109296 2022

[19] [19]

ACL Anthology

Song,X.,Mou,Y.,He,K.,Qiu,Y.,Zhao,J.,Wang,P.,&Xu,W.(2023).ContinualGeneralizedIntentDiscovery:MarchingTowardsDynamic and Open-world Intent Recognition. ACL Anthology. https://doi.org/10.18653/v1/2023.findings-emnlp.289

work page doi:10.18653/v1/2023.findings-emnlp.289 2023

[20] [20]

CLINC150. 2020. UCI Machine Learning Repository. https://doi.org/10.24432/C5MP58. First Author et al.:Preprint submitted to ElsevierPage 19 of 19

work page doi:10.24432/c5mp58 2020