REVIEW 3 major objections 1 minor 1 cited by
ME-TST+: Micro-expression Analysis via Temporal State Transition with ROI Relationship Awareness
T0 review · 3 major / 1 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read This paper claims micro-expression spotting and recognition are better done as one video-level regression over temporal state transitions than as fixed-window classification, and it proposes two Mamba-based architectures, ME-TST and ME-TST+
desk verdict The abstract describes ME-TST+ micro-expression analysis, but the full text is an unrelated 3D world generation paper, so there is nothing to review. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the temporal state transition mechanism inside a state space model (the Mamba-style architecture), which replaces conventional window-level classification with video-level regression. A state transition carries the temporal state of the facial expression forward frame by frame, so the model outputs continuous predictions for onset, apex, and offset rather than classifying fixed windows. ME-TST+ builds on this with multi-granularity ROI relationship modeling and a slowfast Mamba structure; the synergy strategy couples spotting and recognition at both feature and result levels.
What would settle it
Run the released code on a benchmark with micro-expressions of strongly varying duration and compare frame-level spotting F1 and recognition accuracy against the strongest sliding-window baseline; if video-level regression does not beat that baseline on the long-tail durations, the core claim fails.
Extended reading notes
Core claim
The central claim is that temporal state transition mechanisms in state space models can model the onset-apex-offset dynamics of micro-expressions more precisely than fixed-window hard classification, while also supporting micro-expressions of different durations. ME-TST is the baseline design; ME-TST+ adds multi-granularity ROI modeling and a slowfast Mamba framework to limit information loss when treating ME analysis as a time-series task, and a synergy strategy that couples spotting and recognition. The paper asserts this joint formulation outperforms previous approaches. Since the full-text payload is a different manuscript, this statement of the discovery rests on the abstract.
Load-bearing premise
If micro-expression onset, apex, and offset dynamics are not faithfully captured by temporal state transitions in a state space model—that is, if video-level regression on Mamba features is no better than fixed-window classification—the paper's main advantage collapses; the full text provided is an unrelated manuscript, so this premise could not be checked.
Editorial extensions
If this is right
- If the claim holds, micro-expression systems no longer need to pick a window length, because video-level regression can accommodate expressions of varying duration.
- Spotting and recognition can share one model and reinforce each other, instead of being solved as separate tasks.
- Mamba-style state transitions offer a path to long-video micro-expression analysis without the compute cost of windowed classifiers.
- The reported state-of-the-art performance would mean the proposed architectures are the new reference point on the evaluated benchmarks.
Reading between the lines
- If video-level regression on state-space features is indeed the right inductive bias, the same mechanism could transfer to other events with variable onset-apex-offset shapes, such as pain expression, drowsiness, or deception cues, where fixed-window classifiers also struggle.
- The synergy strategy suggests a testable hypothesis: pretraining the spotting and recognition heads jointly on coarsely timestamped video may improve both, even when one label modality is noisy.
- Because the full text supplied with this submission is an unrelated paper, the concrete architecture details and benchmark numbers are unverified; before relying on the claim, a reader should inspect the released code and the paper's actual experimental section.
- A direct comparison of ME-TST+ against its own single-task ablations would reveal how much of the gain comes from the synergy strategy versus the state transition backbone.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The submission, as represented by its abstract, proposes two state-space model architectures for micro-expression (ME) spotting and recognition: ME-TST and ME-TST+. The claimed contributions are temporal-state-transition mechanisms that replace sliding-window classification with video-level regression, multi-granularity ROI modeling, a slowfast Mamba framework, and a synergy strategy linking spotting and recognition at the feature and result levels. The abstract further claims state-of-the-art performance and provides a GitHub link. However, the supplied full text is not this paper: it is 'Matrix-3D: Omnidirectional Explorable 3D World Generation' (arXiv:2508.08086). The body contains no formulation of the state-space models, no description of the ME-TST+ architecture, no micro-expression datasets (e.g., CASME, SAMM), no experimental protocol, and no results for ME spotting or recognition. Consequently, the central contribution described in the abstract is entirely absent from the reviewed document, and none of its load-bearing claims can be inspected or verified.
Significance. If the abstract's claims were supported, the paper would address a genuine and timely problem: the limitations of fixed-window, hard-classification approaches to micro-expression analysis and the artificial separation of spotting and recognition. A unified video-level regression framework built on state-space models with explicit temporal state transitions and ROI relationship awareness could be a meaningful contribution to the field, provided it is accompanied by rigorous evaluation on established ME benchmarks. However, because the submitted full text is an unrelated paper, the proposed method, its derivations, and its empirical evidence are unavailable. No assessment of significance beyond the abstract-level claim is possible. The paper currently offers only a plausibly interesting research direction and a GitHub URL, with no inspectable content to support the claimed state-of-the-art performance.
major comments (3)
- [Abstract vs. Full Text] The abstract describes ME-TST and ME-TST+, but the full text (Sections 1–7 and Appendices A–C) is the unrelated paper 'Matrix-3D: Omnidirectional Explorable 3D World Generation'. None of the components named in the abstract—temporal state transition, slowfast Mamba, multi-granularity ROI modeling, or the spotting/recognition synergy strategy—appear anywhere in the body. The only equation in Section 3 (Eq. 1) is the flow-matching training objective of the Matrix-3D video diffusion model, not a state-space model for micro-expressions. This is not a missing ablation or a presentation gap; the object of review is absent. The central claim of 'state-of-the-art performance' is therefore unsupported by any inspectable evidence.
- [Section 6 (Results)] The Results section reports FID, FVD, rotation error, translation error, PSNR/SSIM/LPIPS, and ablations for panoramic video generation and 3D reconstruction. These metrics and experiments have no bearing on micro-expression spotting or recognition. There is no evaluation protocol for ME-TST+, no benchmark dataset is named, no baselines are compared, and no spotting/recognition accuracy, F1-score, or AUC is reported. The experimental claims in the abstract are thus entirely unverifiable from this document.
- [Whole manuscript] The submission contains no architecture diagram, pseudocode, hyperparameter settings, or model definition for ME-TST/ME-TST+. The only artifact specific to the claimed method is the GitHub repository URL in the abstract. A reviewer cannot check the method's novelty, correctness, reproducibility, or falsifiability. This is a load-bearing defect that cannot be remedied by minor revisions to the current text; the correct manuscript would need to be supplied.
minor comments (1)
- [Abstract] The abstract's GitHub link is not matched by any corresponding code or documentation in the body. If the correct manuscript is resubmitted, the link should be accompanied by a reproducibility statement and a clear mapping to the proposed architectures.
Circularity Check
No circularity identified; the ME-TST+ claims are unverifiable because the submitted full text is an unrelated paper (Matrix-3D).
full rationale
The abstract of arXiv:2508.08082 claims state-of-the-art micro-expression (ME) spotting and recognition via two Mamba-based state-space architectures (ME-TST/ME-TST+), temporal state transition, video-level regression, multi-granularity ROI modeling, slowfast Mamba, and a spotting-recognition synergy strategy. However, the submitted full text is an entirely different paper, 'Matrix-3D: Omnidirectional Explorable 3D World Generation' (arXiv:2508.08086). That body contains no ME-TST architecture, no state-space formulation, no micro-expression datasets (e.g., CASME, SAMM), no experimental protocol, no numerical results, and no ablation for the claimed method. Consequently, there is no derivation chain in front of me whose outputs can be checked against inputs. I cannot exhibit any equation or fitted parameter that reduces to its own input, and there is no load-bearing self-citation chain to audit. The abstract's 'Extensive experiments demonstrate state-of-the-art performance' is an unsupported assertion, not a circular derivation. This is a fatal missing-support/structural defect that makes the submission unverdictable as it stands, but it is not circularity. Under the hard rule not to manufacture circularity, the appropriate circularity score is 0.
Assumptions & free parameters
assumptions (2)
- domain assumption Micro-expression intervals and emotion categories can be jointly recovered by video-level regression in a state space model, without fixed-window priors.
- domain assumption The benchmark datasets used (unnamed in the abstract) provide reliable micro-expression interval and label annotations.
Cite this review
Pith. "Pith review of ME-TST+: Micro-expression Analysis via Temporal State Transition with ROI Relationship Awareness." pith.science (2026). https://pith.science/paper/VOFKQZT2
@misc{pith2026250808082,
author = {Pith},
title = {Pith review of: ME-TST+: Micro-expression Analysis via Temporal State Transition with ROI Relationship Awareness},
year = {2026},
howpublished = {\url{https://pith.science/paper/VOFKQZT2}},
note = {Machine review of arXiv:2508.08082}
}
read the original abstract
Micro-expressions (MEs) are regarded as important indicators of an individual's intrinsic emotions, preferences, and tendencies. ME analysis requires spotting of ME intervals within long video sequences and recognition of their corresponding emotional categories. Previous deep learning approaches commonly employ sliding-window classification networks. However, the use of fixed window lengths and hard classification presents notable limitations in practice. Furthermore, these methods typically treat ME spotting and recognition as two separate tasks, overlooking the essential relationship between them. To address these challenges, this paper proposes two state space model-based architectures, namely ME-TST and ME-TST+, which utilize temporal state transition mechanisms to replace conventional window-level classification with video-level regression. This enables a more precise characterization of the temporal dynamics of MEs and supports the modeling of MEs with varying durations. In ME-TST+, we further introduce multi-granularity ROI modeling and the slowfast Mamba framework to alleviate information loss associated with treating ME analysis as a time-series task. Additionally, we propose a synergy strategy for spotting and recognition at both the feature and result levels, leveraging their intrinsic connection to enhance overall analysis performance. Extensive experiments demonstrate that the proposed methods achieve state-of-the-art performance. The codes are available at https://github.com/zizheng-guo/ME-TST.
Forward citations
Cited by 1 Pith paper
-
Boosting Micro-Expression Analysis via Prior-Guided Video-Level Regression
A prior-guided video-level regression with adaptive interval selection and full parameter sharing sets new state-of-the-art results on micro-expression spotting and recognition benchmarks.
Reference graph
Works this paper leans on
-
[1]
Matrix-3D: Omnidirectional Explorable 3D World Generation Zhongqi Yang 1∗ Wenhang Ge 2∗ Yuqi Li 1,3∗ ∗Jiaqi Chen 1† Haoyuan Li 1† † Mengyin An1 Fei Kang1 Hua Xue1 Baixin Xu1 Yuyang Yin1 Eric Li1 Yang Liu1 Yikai Wang4 Hao-Xiang Guo1‡ ‡Yahui Zhou1 1 Skywork AI 2 Hong Kong University of Science and Technology (Guangzhou) 3 Institute of Computing Technology, ...
arXiv 2025
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.