Pith. sign in

REVIEW 3 major objections 7 minor 3 cited by

Micro-expression action units are detected more accurately when multi-scale visual evidence and anatomical AU-relation tokens are fused into structured prompts that condition a large language model for deduction.

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

2026-07-15 12:39 UTC pith:JC3TGMGM

load-bearing objection Solid engineering extension of AULLM that delivers real LOSO and cross-domain Macro-F1 gains on three ME-AU sets; the FACS-prior story is a bit oversold but the tables and ablations still hold. the 3 major comments →

arxiv 2603.08387 v2 pith:JC3TGMGM submitted 2026-03-09 cs.CV

AULLM++: Structured-Token-Conditioned Large Language Models for Micro-Expression Action Unit Detection

Zhishu Liu , Kaishen Yuan , Bo Zhao , Hui Ma , Zitong Yu This is my paper
classification cs.CV
keywords Micro-ExpressionAction Unit DetectionLarge Language ModelGraph Neural NetworkFacial Action Coding SystemCounterfactual RegularizationCross-Domain Generalization
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.

Micro-expressions produce only fleeting, low-intensity muscle signals that conventional vision models easily lose to noise, coarseness, and ignored muscle co-dependencies. This paper claims that the right remedy is not heavier visual pooling but a three-stage reasoning pipeline: build a compact multi-granularity content token that preserves high-frequency texture alongside global context, encode FACS-style anatomical synergies and inhibitions as an instance-adaptive instruction token, then let a large language model deduce multi-label AU presence from the joint structured prompt. A training-only counterfactual intervention on the instruction tokens further severs dataset-specific shortcuts. If the claim holds, affective computing gains a more reliable, generalizable foundation for reading involuntary facial cues that pure pattern matching has long mishandled.

Core claim

AULLM++ reaches state-of-the-art micro-expression AU detection (CASME II 82.4 %, SAMM 62.6 %, 4DME-Micro 57.7 % Macro-F1 under LOSO; 51.9 % average cross-domain) by converting multi-granularity visual features into a content token and FACS-prior AU relations into an instruction token, conditioning an LLM on their fusion, and applying counterfactual consistency regularization only at training time.

What carries the argument

The structured prompt formed by the Content Token (MGE-EFP gated fusion of Laplacian-enhanced mid-level textures with high-level semantics) and the Instruction Token (R-AUGNN sparse FACS prior plus instance-adaptive graph message passing), which together condition LoRA-adapted LLM deduction; Counterfactual Consistency Regularization intervenes only on instruction tokens during training.

Load-bearing premise

The hand-specified sparse anatomical prior among action units, once fused with instance attention, supplies genuine domain-invariant causal structure that legitimately guides the language model rather than merely encoding co-occurrence statistics of the training sets.

What would settle it

On a held-out micro-expression corpus whose AU co-occurrence statistics systematically violate the FACS prior, or in a full ablation that replaces the prior graph with a fully data-driven or empty graph, Macro-F1 and cross-domain transfer fall to or below pure visual baselines; that outcome would falsify the claim that the structural prior supplies useful causal guidance.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Micro-expression AU detectors can move from fragile visual pattern matching to guided logical deduction and thereby raise both within-domain Macro-F1 and cross-dataset transfer.
  • Complex, visually similar AU combinations become more separable once synergistic and inhibitory muscle relationships are supplied as explicit instructions.
  • Cross-domain robustness can be improved without any inference-time cost by intervening on structural tokens only during training.
  • Affective systems that need fine-grained involuntary cues gain a concrete architecture that anchors language-model reasoning on physical facial anatomy.

Where Pith is reading between the lines

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

  • The same content-plus-instruction token pattern could transfer to other low-signal multi-label problems such as subtle medical video events or rare industrial anomalies.
  • If the LLM is functioning mainly as a soft multi-label decoder, lighter reasoning heads conditioned on the same tokens may deliver most of the gain at lower compute cost.
  • Transparent step-by-step AU deduction, already flagged as future work, would make the same pipeline auditable for high-stakes behavioral analysis.
  • Learning or adapting the anatomical prior from large macro-expression corpora rather than fixing it by hand could further reduce residual domain gaps.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. AULLM++ reformulates micro-expression Action Unit (AU) detection as a three-stage process—evidence construction, structure modeling, and LLM-based deduction—rather than pure visual multi-label classification. A Multi-Granularity Evidence-Enhanced Fusion Projector (MGE-EFP) fuses Laplacian-enhanced mid-level textures with high-level semantics into a compact Content Token T_v (Eqs. 1–4). A Relation-Aware AU Graph Neural Network (R-AUGNN) combines a hand-specified sparse FACS/psychological prior adjacency A_prior with instance-adaptive attention into  (Eq. 6), then GCN-encodes node features into Instruction Tokens τ_au. These tokens are concatenated into a structured LLM prompt (DeepSeek-R1-Distill-Qwen-1.5B with LoRA); training-only Counterfactual Consistency Regularization (CCR, Eq. 8) intervenes on instruction tokens to encourage target-AU flip and non-target invariance. On LOSO, the method reports Macro-F1 of 82.4% (CASME II), 62.6% (SAMM), and 57.7% (4DME-Micro), with cross-domain average 51.9% over six transfer directions, outperforming the conference AULLM and LED-based baselines. Ablations (Table V) and t-SNE/attention visualizations support each module.

Significance. If the empirical gains hold under stronger statistical scrutiny and a fully disclosed prior, the work is a meaningful step for micro-expression AU detection: it shows that highly compressed multi-granularity visual tokens plus explicit AU-relation tokens can condition a small LLM more effectively than dense visual pooling or motion-magnification CNNs, especially under cross-dataset shift. Strengths include a clear extension path from AULLM, systematic LOSO and six-way transfer evaluation, module ablations that isolate R-AUGNN/MGE-EFP/CCR/LLM, and a training-only regularizer with no inference cost. The framing of micro-expression AU detection as structured-token-conditioned deduction is useful for the community even if the LLM is primarily acting as a strong multi-label head rather than producing explicit chain-of-thought traces. Code release upon acceptance would further raise impact.

major comments (3)
  1. Methodology, R-AUGNN / Eq. (6): A_prior is described only as “pre-construct[ed] … based on extensive psychological statistics and FACS anatomical rules” and is never listed, derived, or released. The paper’s central narrative attributes cross-domain robustness to “domain-invariant physical rules” and “anatomical physical priors” that guide LLM deduction. Without the actual sparse adjacency (which edges, which weights, how “psychological statistics” were separated from dataset co-occurrence), the claim that R-AUGNN supplies causal anatomical structure rather than transferable co-occurrence bias cannot be audited. Table V shows FACS graph > fully-connected/self-loops, but that does not identify the source of the edges. Please publish A_prior (or a reproducible construction procedure independent of the evaluation sets) and discuss residual risk that non-zero edges encode long-tail co-occurr
  2. §IV.B–D and Tables I–IV: All headline Macro-F1 numbers are single-run point estimates. Datasets are tiny and long-tailed (CASME II 247 clips / 26 subjects; SAMM 159 / 32), yet training uses batch size 256 for 350 epochs with differential LRs and several free coefficients (α, γ, λ_ccr, λ_inv, λ_Δ). Within-domain gains over AULLM are small (CASME II 81.4→82.4; SAMM 61.9→62.6). Without multi-seed means±std, or at least LOSO fold variance / significance tests against AULLM and SSSNet LED, the SOTA and “superior cross-domain generalization” claims are under-supported. Please report multi-seed or fold-level variance for the main tables and clarify early-stopping / selection protocol under LOSO.
  3. §III, LLM-Driven Deduction and CCR (Eq. 8, Algorithm 1): The manuscript repeatedly frames the LLM stage as “logical deduction,” “causal deduction,” and “reasoning,” but the implemented pipeline freezes the LLM, injects LoRA, concatenates T_v and τ_au, and applies a linear multi-label head on a single hidden state—no intermediate reasoning traces, constrained decoding, or AU-consistency constraints at generation time. CCR forces target flip and non-target KL invariance under hand-crafted instruction perturbations; that is a useful regularizer, but calling it isolation of “authentic causal relationships” overstates what is demonstrated. Please either (i) provide evidence that the LLM performs non-trivial multi-AU logical inference beyond a strong classifier, or (ii) tone the causal/reasoning language to match the actual mechanism (structured-token-conditioned multi-label prediction with in
minor comments (7)
  1. Title inconsistency: the manuscript header uses “Structural Reasoning … for Micro-Expression Recognition,” while the abstract/arXiv-style title emphasizes “Structured-Token-Conditioned … Action Unit Detection.” Align title with the actual task (AU detection).
  2. Fig. 2 / Algorithm 1: CCR is said to loop over each AU k with re-forwarding; aggregation of L_ccr across k (sum vs mean) and interaction with batch size 256 should be stated explicitly for reproducibility.
  3. Notation: Content Token is T_v in the body but CT in the abstract; Instruction Token is τ_au / IT. Unify early and keep one symbol set.
  4. Related Work cites “[?], [26]” with a missing reference marker; fix the broken citation.
  5. Table III AU set (includes AU6, AU45) differs from CASME II/SAMM; briefly justify AU selection per dataset so Macro-F1 averages are comparable in spirit.
  6. Fig. 3 and Fig. 4 captions are thin; add dataset/AU labels and what “Baseline” is in Fig. 4 so qualitative claims are self-contained.
  7. Extensions paragraph claims “far more extensive experiments” including 4DME-Micro; ensure conference AULLM numbers on 4DME/SAMM are from the same protocol as this paper to avoid apples-to-oranges comparison.

Circularity Check

1 steps flagged

No derivation-by-construction circularity: SOTA and cross-domain claims are empirical Macro-F1 on held-out subjects/datasets; only mild self-extension of the authors' conference AULLM baseline.

specific steps
  1. self citation load bearing [Extensions from Conference Version; Tables I–IV (AULLM Conf. baseline)]
    "This paper is a substantially extended version of our prior preliminary conference publication, AULLM [15]. ... AULLM (Conf.) [15] 92.2 88.4 ... 81.4 / AULLM++ (Ours) 93.1 ... 82.4 ... AULLM (Conf.) [15] ... 47.3 / AULLM++ (Ours) ... 51.9"

    The narrative of evolutionary gains (structural priors, MGE-EFP, CCR) is partly framed against the authors' own conference system. This is ordinary self-baseline comparison, not load-bearing uniqueness: SOTA claims also rest on independent external methods (LED, SCA, DVASP, ResNet, LBP-TOP) and held-out LOSO/cross-domain metrics. Mild only; does not make test Macro-F1 tautological.

full rationale

AULLM++ is an empirical multi-label detection framework, not a first-principles derivation. Content tokens (MGE-EFP), instruction tokens (R-AUGNN with hand-specified sparse A_prior fused as Â=α A_prior+(1−α)A_dynamic), LLM LoRA decoding, and training-only CCR (flip target AU / KL-invariance on others) are architectural and regularization choices. Reported results (CASME II 82.4%, SAMM 62.6%, 4DME-Micro 57.7%; cross-domain avg 51.9%) are Macro-F1 under LOSO and train→test transfer against external baselines and the authors' prior conference system; they are not algebraic identities of the inputs. A_prior is an author-constructed FACS/psychological ansatz (not fitted to the test labels and then re-predicted), α and GCN weights are learned on training folds only, and CCR is disabled at inference. Self-citation of AULLM [15] is normal extended-work framing and is used as a comparable baseline, not as a uniqueness theorem that forces the present claims. No step reduces a claimed prediction to its own definition or fit by construction. Residual concerns about whether the hand-specified prior encodes co-occurrence statistics rather than pure anatomy are assumption/correctness issues, not circularity under the stated criteria.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central empirical claim rests on standard multi-label ME evaluation practice plus several paper-specific modeling choices: a hand-built sparse AU prior, multi-scale fusion hyperparameters, LoRA-adapted LLM decoding, and a training-only counterfactual intervention loss. Free parameters are the usual deep-learning knobs plus explicit balancing coefficients (α, γ, λ_ccr, etc.). Invented entities are architectural modules, not new physical objects; they lack independent evidence outside the reported benchmarks.

free parameters (5)
  • α (static prior vs dynamic attention mix in Â)
    Learnable balance between FACS prior adjacency and instance attention; directly shapes instruction tokens that condition the LLM.
  • γ (Laplacian high-frequency enhancement strength)
    Controls mid-level texture boost in MGE-EFP before gated fusion into the content token.
  • λ_ccr, λ_inv, λ_Δ (CCR loss weights)
    Empirically set regularizer strengths (λ_ccr=0.1 reported) that force flip/invariance under instruction perturbations during training.
  • LoRA rank r=16, scale α=32; LR 2e-4/1e-4; batch 256; 350 epochs
    Optimization and adapter hyperparameters that materially affect convergence on small long-tail ME datasets.
  • GCN depth L=2 and AU node feature dimension
    Graph capacity choices that determine how multi-hop AU relations are encoded into τ_au.
axioms (5)
  • domain assumption FACS anatomical/psychological rules define a useful sparse prior over AU synergy and inhibition for micro-expressions.
    Invoked to build A_prior in R-AUGNN; ablations treat this prior as load-bearing versus fully-connected or self-loop graphs.
  • domain assumption Micro-expression AU detection is multi-label binary classification under LOSO with Macro-F1 as the primary metric.
    Standard community protocol stated in Experiments; underpins all SOTA comparisons.
  • ad hoc to paper Compressing fused visual features and graph instructions into a few tokens preserves enough signal for LLM multi-label deduction without dense patch flooding.
    Core design premise of the content/instruction token interface versus standard VLM patch prompting.
  • ad hoc to paper Instruction-level counterfactual flips isolate causal AU logic better than pixel-space interventions for domain robustness.
    Justification for CCR in Methodology and Related Work; training-only, no inference cost.
  • standard math Standard GCN message passing and BCE multi-label loss are valid computational substrates.
    Used without modification in Eqs. for H^(l+1) and L_cls.
invented entities (3)
  • MGE-EFP Content Token (CT / T_v) no independent evidence
    purpose: Distill mid-level high-frequency textures and high-level semantics into a compact LLM-aligned visual premise.
    Architectural construct; evidence is internal ablations and attention maps, not an external measurable object.
  • R-AUGNN Instruction Token (IT / τ_au) no independent evidence
    purpose: Encode sparse FACS prior plus instance-adaptive AU interactions as structural guidance for the LLM.
    Paper-specific graph-to-token interface; validated only via within-paper ablations and transfer gains.
  • Counterfactual Consistency Regularization (CCR) on instruction tokens no independent evidence
    purpose: Training intervention that flips a target AU instruction and enforces prediction flip plus non-target invariance.
    Regularizer invented for this pipeline; no external causal benchmark beyond reported Macro-F1 deltas.

reviewed 2026-07-15 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AULLM++: Structured-Token-Conditioned Large Language Models for Micro-Expression Action Unit Detection." pith.science (2026). https://pith.science/paper/JC3TGMGM

@misc{pith2026260308387,
  author       = {Pith},
  title        = {Pith review of: AULLM++: Structured-Token-Conditioned Large Language Models for Micro-Expression Action Unit Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JC3TGMGM}},
  note         = {Machine review of arXiv:2603.08387}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Micro-expression Action Unit (AU) detection identifies localized AUs from subtle facial muscle activations, providing a foundation for decoding affective cues. Previous methods face three key limitations: (1) heavy reliance on low-density visual information, rendering discriminative evidence vulnerable to background noise; (2) coarse-grained feature processing that misaligns with the demand for fine-grained representations; and (3) neglect of inter-AU correlations, restricting the parsing of complex expression patterns. We propose AULLM++, a reasoning-oriented framework leveraging Large Language Models (LLMs), which injects visual features into textual prompts as actionable semantic premises to guide inference. It formulates AU prediction into three stages: evidence construction, structure modeling, and deduction-based prediction. Specifically, a Multi-Granularity Evidence-Enhanced Fusion Projector (MGE-EFP) fuses mid-level texture cues with high-level semantics, distilling them into a compact Content Token (CT). Furthermore, inspired by micro- and macro-expression AU correspondence, we encode AU relationships as a sparse structural prior and learn interaction strengths via a Relation-Aware AU Graph Neural Network (R-AUGNN), producing an Instruction Token (IT). We then fuse CT and IT into a structured textual prompt and introduce Counterfactual Consistency Regularization (CCR) to construct counterfactual samples, enhancing the model's generalization. Extensive experiments demonstrate AULLM++ achieves state-of-the-art performance on standard benchmarks and exhibits superior cross-domain generalization.

Figures

Figures reproduced from arXiv: 2603.08387 by Bo Zhao, Hui Ma, Kaishen Yuan, Zhishu Liu, Zitong Yu.

Figure 1
Figure 1. Figure 1: (a) Micro-expression AU detection is challenged [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of AULLM++. The model constructs a compact visual evidence token [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: t-SNE visualization of high-level features across [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of feature evolution. priors from R-AUGNN and the causal intervention of CCR, our framework successfully forces the model to discard a significant portion of dataset-specific environmental pseudo￾correlations. Consequently, features from different datasets are projected into a much closer, generalized semantic space. While perfectly aligning distinct datasets into a singular, domain-agnostic … view at source ↗
Figure 3
Figure 3. Figure 3: In the standard baseline, features exhibit severe domain [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. Navigating the Emotion Tree: Hierarchical Hyperbolic RAG for Multimodal Emotion Recognition

    cs.LG 2026-05 unverdicted novelty 6.0

    HyperEmo-RAG uses hierarchical hyperbolic embeddings and graph-based evidence injection to outperform prior methods in multimodal emotion recognition.

  2. GMoT: Gated Motion-Aware Tokenization for Fine-Grained Micro-Gesture Video Reasoning with Multimodal LLMs

    cs.CV 2026-07 conditional novelty 5.0

    GMoT's gated motion tokens improve multimodal LLM micro-gesture recognition on iMiGUE and SMG, with limited support for reasoning-grounding claims.

  3. AffectAgent: Collaborative Multi-Agent Reasoning for Retrieval-Augmented Multimodal Emotion Recognition

    cs.CV 2026-04 unverdicted novelty 5.0

    AffectAgent deploys a query planner, evidence filter, and emotion generator as collaborative agents trained via MAPPO with shared reward, plus MB-MoE and RAAF modules, to achieve superior multimodal emotion recognitio...

Reference graph

Works this paper leans on

45 extracted references · 10 linked inside Pith · cited by 3 Pith papers

  1. [1]

    Facial action coding system: a technique for the measurement of facial movement,

    P. Ekman and W. V . Friesen, “Facial action coding system: a technique for the measurement of facial movement,”Consulting Psychologists Press, 1978

    1978

  2. [2]

    Auformer: Vision transformers are parameter-efficient facial action unit detectors,

    K. Yuan, Z. Yu, X. Liu, W. Xie, H. Yue, and J. Yang, “Auformer: Vision transformers are parameter-efficient facial action unit detectors,” inECCV, 2024

    2024

  3. [3]

    Automatic analysis of facial actions: A survey,

    B. Martinez, M. F. Valstar, B. Jiang, and M. Pantic, “Automatic analysis of facial actions: A survey,”IEEE Transactions on Affective Computing, 2019

    2019

  4. [4]

    Representation flow for action recognition,

    A. Piergiovanni and M. S. Ryoo, “Representation flow for action recognition,” inCVPR, 2019

    2019

  5. [5]

    Dynamic texture recognition using local binary patterns with an application to facial expressions,

    G. Zhao and M. Pietikainen, “Dynamic texture recognition using local binary patterns with an application to facial expressions,”IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2007

    2007

  6. [6]

    Learning spatiotemporal features with 3d convolutional networks,

    D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri, “Learning spatiotemporal features with 3d convolutional networks,”

  7. [7]

    Available: https://arxiv.org/abs/1412.0767

    [Online]. Available: https://arxiv.org/abs/1412.0767

    Pith/arXiv arXiv

  8. [8]

    Spon- taneous facial micro-expression recognition using 3d spatiotemporal convolutional neural networks,

    S. P. Teja Reddy, S. Teja Karri, S. R. Dubey, and S. Mukherjee, “Spon- taneous facial micro-expression recognition using 3d spatiotemporal convolutional neural networks,” in2019 International Joint Conference on Neural Networks (IJCNN), 2019

    2019

  9. [9]

    A main directional mean optical flow feature for spontaneous micro- expression recognition,

    Y .-J. Liu, J.-K. Zhang, W.-J. Yan, S.-J. Wang, G. Zhao, and X. Fu, “A main directional mean optical flow feature for spontaneous micro- expression recognition,”IEEE Transactions on Affective Computing, 2016

    2016

  10. [10]

    Facial action unit detection with local key facial sub-region based multi-label classification for micro-expression analysis,

    L. Zhang, O. Arandjelovic, and X. Hong, “Facial action unit detection with local key facial sub-region based multi-label classification for micro-expression analysis,” inProceedings of the 1st Workshop on Facial Micro-Expression: Advanced Techniques for Facial Expressions Generation and Spotting, 2021

    2021

  11. [11]

    Semantic relationships guided representation learning for facial action unit recognition,

    G. Li, X. Zhu, Y . Zeng, Q. Wang, and L. Lin, “Semantic relationships guided representation learning for facial action unit recognition,” in AAAI, 2019

    2019

  12. [12]

    Au-assisted graph attention convolutional network for micro-expression recognition,

    H.-X. Xie, L. Lo, H.-H. Shuai, and W.-H. Cheng, “Au-assisted graph attention convolutional network for micro-expression recognition,” in ACM MM, 2020

    2020

  13. [13]

    Large language models: A survey,

    S. Minaee, T. Mikolov, N. Nikzad, M. Chenaghlu, R. Socher, X. Am- atriain, and J. Gao, “Large language models: A survey,”arXiv preprint arXiv:2402.06196, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2301.12597

    Pith/arXiv arXiv 2023

  15. [15]

    Coun- terfactual vqa: A cause-effect look at language bias,

    Y . Niu, K. Tang, H. Zhang, Z. Lu, X.-S. Hua, and J.-R. Wen, “Coun- terfactual vqa: A cause-effect look at language bias,” inCVPR, 2021

    2021

  16. [16]

    Au-llm: Micro-expression action unit detection via enhanced llm-based feature fusion,

    Z. Liu, K. Yuan, B. Zhao, Y . Xu, and Z. Yu, “Au-llm: Micro-expression action unit detection via enhanced llm-based feature fusion,” inChinese Conference on Biometric Recognition (CCBR), 2026

    2026

  17. [17]

    4dme: A spontaneous 4d micro-expression dataset with multimodalities,

    X. Li, S. Cheng, Y . Li, M. Behzad, J. Shen, S. Zafeiriou, M. Pantic, and G. Zhao, “4dme: A spontaneous 4d micro-expression dataset with multimodalities,”IEEE Transactions on Affective Computing, 2023

    2023

  18. [18]

    Micro-expression action unit detection with spatial and channel attention,

    Y . Li, X. Huang, and G. Zhao, “Micro-expression action unit detection with spatial and channel attention,”Neurocomputing, 2021

    2021

  19. [19]

    Micro-expression action unit detection with dual-view attentive similarity-preserving knowledge distillation,

    Y . Li, W. Peng, and G. Zhao, “Micro-expression action unit detection with dual-view attentive similarity-preserving knowledge distillation,” in2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), 2021

    2021

  20. [20]

    Learnable eulerian dynamics for micro-expression action unit detection,

    T. Varanka, W. Peng, and G. Zhao, “Learnable eulerian dynamics for micro-expression action unit detection,” inImage Analysis, 2024

    2024

  21. [21]

    Infused suppression of magnification artefacts for micro-au detection,

    H.-Q. Khor, Y . Li, X. Jiang, and G. Zhao, “Infused suppression of magnification artefacts for micro-au detection,” inSCIA, 2025

    2025

  22. [22]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inCVPR, 2016

    2016

  23. [23]

    Uniemo: Unifying emotional understanding and generation with learnable expert queries,

    Y . Zhu, L. Zhang, Z. Yu, R. Shao, T. Tan, and L. Nie, “Uniemo: Unifying emotional understanding and generation with learnable expert queries,” arXiv preprint arXiv:2507.23372, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Joint patch and multi-label learning for facial action unit detection,

    K. Zhao, W.-S. Chu, F. De la Torre, J. F. Cohn, and H. Zhang, “Joint patch and multi-label learning for facial action unit detection,” inCVPR, 2015

    2015

  25. [26]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” Advances in neural information processing systems, 2023

    2023

  26. [27]

    Emosym: A symbiotic framework for unified emotional understanding and generation via latent reasoning,

    Y . Zhu, Y . Lyu, Z. Yu, R. Shao, K. Zhou, and L. Nie, “Emosym: A symbiotic framework for unified emotional understanding and generation via latent reasoning,” inProceedings of the 33nd ACM International Conference on Multimedia, 2025

    2025

  27. [28]

    Domain adversarial neural networks for domain generalization: When it works and how to improve,

    A. Sicilia, X. Zhao, and S. J. Hwang, “Domain adversarial neural networks for domain generalization: When it works and how to improve,” 2022. [Online]. Available: https://arxiv.org/abs/2102.03924

    Pith/arXiv arXiv 2022

  28. [29]

    The graph neural network model,

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini, “The graph neural network model,”IEEE Transactions on Neural Networks, 2009

    2009

  29. [30]

    Counterfactual explanations and algorithmic recourses for machine learning: A review,

    S. Verma, V . Boonsanong, M. Hoang, K. E. Hines, J. P. Dickerson, and C. Shah, “Counterfactual explanations and algorithmic recourses for machine learning: A review,” 2022. [Online]. Available: https: //arxiv.org/abs/2010.10596

    Pith/arXiv arXiv 2022

  30. [31]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in CVPR, 2018

    2018

  31. [32]

    Temporal and spatial feature fusion framework for dynamic micro expression recognition,

    F. Liu, B. Nan, X. Qian, and X. Fu, “Temporal and spatial feature fusion framework for dynamic micro expression recognition,”arXiv preprint arXiv:2505.16372, 2025

    Pith/arXiv arXiv 2025

  32. [33]

    Constrained and directional ensemble attention for facial action unit detection,

    Z. Shao, B. Chen, Y . Zhou, X. Shi, C. Li, L. Ma, and D.-Y . Yeung, “Constrained and directional ensemble attention for facial action unit detection,”Pattern Recognition, 2026

    2026

  33. [34]

    Micro-expression classification based on landmark relations with graph attention convolutional net- work,

    A. J. Rakesh Kumar and B. Bhanu, “Micro-expression classification based on landmark relations with graph attention convolutional net- work,” in2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2021

    2021

  34. [35]

    Visual prompt tuning,

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” inEuropean conference on computer vision, 2022

    2022

  35. [36]

    Learning to prompt for vision- language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Learning to prompt for vision- language models,”IJCV, 2022

    2022

  36. [37]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models,” inICLR, 2022

    2022

  37. [38]

    Fitnets: Hints for thin deep nets,

    A. Romero, N. Ballas, S. E. Kahou, A. Chassang, G. Gkioxari, and Y . Bengio, “Fitnets: Hints for thin deep nets,”arXiv preprint arXiv:1412.6550, 2014

    Pith/arXiv arXiv 2014

  38. [39]

    Similarity-preserving knowledge distillation,

    F. Tung and G. Mori, “Similarity-preserving knowledge distillation,” in ICCV, 2017

    2017

  39. [40]

    Paying more attention to attention: Improving the performance of convolutional neural networks via atten- tion transfer,

    S. Zagoruyko and N. Komodakis, “Paying more attention to attention: Improving the performance of convolutional neural networks via atten- tion transfer,”arXiv preprint arXiv:1612.03928, 2016

    Pith/arXiv arXiv 2016

  40. [41]

    Casme ii: An improved spontaneous micro-expression database and the baseline evaluation,

    W.-J. Yan, X. Li, S.-J. Wang, G. Zhaoet al., “Casme ii: An improved spontaneous micro-expression database and the baseline evaluation,” PLoS One, 2014

    2014

  41. [42]

    Samm: A spontaneous micro-facial movement dataset,

    A. K. Davison, C. Lansley, N. Costen, K. Tan, and M. H. Yap, “Samm: A spontaneous micro-facial movement dataset,”IEEE Transactions on Affective Computing, 2018

    2018

  42. [43]

    Data leakage and evaluation issues in micro-expression analysis,

    T. Varanka, Y . Li, W. Peng, and G. Zhao, “Data leakage and evaluation issues in micro-expression analysis,”IEEE TAFFC, 2023

    2023

  43. [44]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    DeepSeek-AI, “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,” 2025. [Online]. Available: https: //arxiv.org/abs/2501.12948

    Pith/arXiv arXiv 2025

  44. [45]

    Dire: Enhancing facial expression recognition through domain-invariant representation learning for robust generalization,

    H. Kim, Y . Jung, and B. C. Song, “Dire: Enhancing facial expression recognition through domain-invariant representation learning for robust generalization,”IEEE Transactions on Multimedia, 2025

    2025

  45. [46]

    Micro-expression recognition with super- vised contrastive learning,

    R. Zhi, J. Hu, and F. Wan, “Micro-expression recognition with super- vised contrastive learning,”Pattern Recognition Letters, 2022

    2022

This paper was first reviewed by grok-4.5 on July 15, 2026.