Pith. sign in

REVIEW 4 major objections 6 minor 82 references

FAS-R1: A Unified Multi-Task MLLM for Reasoning Face Anti-Spoofing

T0 review · 4 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A single 3B multimodal language model can match discriminative face-anti-spoofing accuracy while also producing inspectable, image-grounded rationales.

desk verdict Solid RL-based FAS system with clean ablations, but the image-level in-domain split leaks identity and inflates the headline numbers; the rationale evaluation is self-referential, so the central claims need a fix before they can be trusted. read the letter →

arxiv 2607.26432 v1 pith:55LRRMEZ submitted 2026-07-29 cs.CV cs.AI

Hongyang Wang , Yichen Shi , Hongrui Li , Yiru Huo , Jun Feng , Zitong Yu This is my paper
classification cs.CVcs.AI
keywords faceanti-spoofingpresentationattackdetectionmultimodallargelanguagemodelchain-of-thoughtreasoningreinforcementlearningGRPOdomaingeneralizationspoof-regionlocalization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Face anti-spoofing has mostly been a label-only 'real or spoof' classification problem, and explainable variants have tended to sacrifice either accuracy or genuine image grounding. FAS-R1 claims to close that gap with a multimodal large language model that, from a single interface, judges authenticity, names the attack type, draws the spoof region, and writes an image-grounded rationale. Training is two-stage: a cold-start supervised fine-tune on FAS-R1-23K, a curated 23K-sample chain-of-thought dataset, then a face-specific reinforcement-learning stage. Two mechanisms do the work there: Degradation-Simulated Augmentation (DSA) pairs clean and degraded views of the same image inside one rollout group so spoof cues must be quality-stable, and Difficulty-Aware GRPO (DA-GRPO) reweights task–attack subgroups so hard cases are not neglected. On in-domain tests the 3B model reaches 98.75% authenticity accuracy, 93.33% attack-type accuracy, and 96.30/94.73% AP@40/AP@50, and it outperforms compared systems on cross-domain authenticity and answer-and-rationale quality; the 7B version scales further.

What carries the argument

Key machinery: (1) FAS-R1-23K — a 22,996-sample dataset pairing authenticity labels, attack categories, manual spoof-region boxes, and long chain-of-thought rationales in an annotation-constrained <think>/<answer> format, generated by one vision model, verified by another, and rule/manual filtered; (2) FAS-specific GRPO — group-relative policy optimization whose rollout groups mix clean and synthetically degraded views of the same image (DSA) so spoof cues must be quality-stable, and whose advantages are rescaled by an EMA task–attack subgroup proficiency (DA-GRPO) so easy subgroups do not dominate. Rewards cover format, correctness, localization IoU, and rationale–answer consistency.

What would settle it

Run the same training and evaluation protocol using the original released checkpoints of the two compared MLLM baselines (instead of the re-implementations), and check whether FAS-R1's accuracy, localization, and Elo-rationale gaps persist. A second check: ablate DA-GRPO on a held-out hard subgroup (e.g., makeup/mask attacks) and see whether the large coarse-ACC gain from 94.45% to 98.75% reproduces; if it does not, the difficulty-reweighting claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a multimodal LLM can match discriminative face-anti-spoofing accuracy while also naming the attack type, drawing the spoof region, and giving a rationale. The recipe: cold-start SFT on a 23K long-CoT dataset, then a GRPO stage with degraded-pair rollouts (DSA) and task–attack subgroup reweighting (DA-GRPO). Ablations show DA-GRPO is the largest lever, lifting coarse accuracy from 94.45% to 98.75% and cutting HTER to 1.17%; the 7B version improves further.

Load-bearing premise

The reported advantage over the two prior MLLM-based FAS methods rests on the authors' re-implementations of those baselines, because the original code and checkpoints are not public; if those re-implementations are undertuned or misaligned with the new training data, the performance gap could be overstated.

Editorial extensions

If this is right

  • Security deployments can get a binary decision and an inspectable explanation from a single model, without the accuracy loss that has previously separated discriminative and explainable FAS.
  • DSA's paired clean/degraded rollouts improve all five reported metrics over plain GRPO, so quality-stable evidence is learnable from on-policy augmentation rather than from extra augmentation models.
  • DA-GRPO produces the biggest single accuracy gain (94.45% → 98.75% coarse ACC), indicating that easy-sample dominance, not capacity, was the main bottleneck on subtle attacks like makeup and mask.
  • The SFT+RL recipe transfers to larger backbones: 7B improves every in-domain metric and most cross-domain results, so the approach is scalable.
  • Cross-domain authenticity results are competitive with or better than compared methods, suggesting the reasoning supervision does not simply memorize training-domain artifacts.

Reading between the lines

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

  • A direct testable extension is to apply DSA to other vision tasks where outputs must survive sensor-quality shifts (deepfake detection, medical imaging); the mechanism is not face-specific.
  • DA-GRPO's subgroup proficiency maps could be adapted to any multi-task generative model with known semantic classes; the paper leaves open the optimal grouping granularity and reweighting schedule.
  • Because FAS-R1-23K is constructed with commercial generative and verifying models, an open-source replication of the data pipeline would clarify how much of the gain is the dataset versus the algorithm.
  • If the rationale-quality judgments hold, a natural next experiment is a human user study measuring whether reviewers catch spoofs faster or more reliably with FAS-R1's rationales than with a label-only system; the paper does not run that study.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. FAS-R1 proposes a two-stage training framework for a 3B/7B multimodal LLM (Qwen2.5-VL) that unifies face anti-spoofing authenticity classification, attack-type recognition, spoof-region localization, and generated rationales. Stage 1 performs cold-start SFT on a newly constructed 22,996-sample long-CoT corpus (FAS-R1-23K) built from WMCA, PADISI-Face, and SiW-Mv2; Stage 2 applies GRPO with Degradation-Simulated Augmentation (paired clean/degraded rollouts) and Difficulty-Aware GRPO (task–attack subgroup advantage reweighting). The paper reports in-domain coarse ACC 98.75%, fine-grained ACC 93.33%, AP@40/AP@50 of 96.30/94.73%, several cross-domain authenticity generalization results, and higher judged rationale quality than compared systems, concluding that an MLLM can match discriminative FAS accuracy while providing inspectable semantics.

Significance. The task and proposed interface are timely: moving FAS from label-only output to evidence-grounded, multi-task MLLM prediction is a natural next step, and the two-stage SFT-plus-FAS-specific-GRPO recipe is a reasonable engineering contribution with a clear motivation. The paper scores well on internal consistency: the central classification metrics are grounded in external benchmark labels rather than generated labels, the ablation study is controlled (same rewards, prompts, and verifier across RL variants), and the cross-domain protocols are standard. If the reported numbers held up under a stricter evaluation, the paper would make a useful empirical contribution. However, as it stands, the headline in-domain claims are compromised by a subject-overlap protocol, and the comparison to prior MLLM FAS methods rests on unreleased re-implementations. The rationale-quality conclusions also rely largely on self-referential judge models.

major comments (4)
  1. [Evaluation Protocols; Table 2] The in-domain split is described as 'the merged images are split at the image level into training, validation, and test sets with an 8:1:1 ratio.' WMCA, PADISI-Face, and SiW-Mv2 each contain multiple images per subject, so an image-level random split places the same identities in both training and test sets. This allows the model to exploit identity-specific appearance (glasses, skin texture, lighting) rather than generalizable spoof cues, and it directly inflates the headline numbers in Table 2 (98.75% coarse ACC, 93.33% attack ACC, 96.30/94.73 AP@40/AP@50) and the comparison to FaceShield/PA-FAS. The cross-domain results in Table 3 are less affected because domains are disjoint. The authors should re-evaluate with a subject-disjoint split, or at minimum report the train/test identity overlap and the corresponding metrics. Without this, the abstract's in-domain claims are not trustworth
  2. [Implementation Details; Tables 2–3] The paper states that since other MLLM-based FAS methods have not released code/checkpoints, FaceShield and PA-FAS are trained with FAS-R1-23K. The head-to-head comparisons in Tables 2 and 3 therefore depend entirely on how faithfully and how well those baselines are re-implemented, but the paper gives no hyperparameters, training lengths, or validation-based checkpoint selection for the re-implemented baselines, and no sensitivity analysis. A suboptimal re-implementation would make the reported improvements over previous MLLM methods overstated even if FAS-R1 works. Please provide full training details for the re-implementations, an ablation of their training budget, or comparison against published numbers where available, and ideally release the baseline checkpoints to make the comparison reproducible.
  3. [Answer-and-Rationale Quality; Table 4, Fig. 5] The rationale-quality evaluation is not independent of the training pipeline. Training rationales in FAS-R1-23K were generated by Gemini 2.5 Flash and verified by GPT-5; the judges in Table 4/Fig. 5 are Claude-Sonnet-4.5 and Gemini-3-Pro, and the judge models themselves are also included as candidate systems. This creates a risk of style-matching bias: a judge may prefer outputs that resemble its own generation style or the style of the same model family. The claim that FAS-R1 produces better answer-and-rationale quality is a central contribution, so the authors should add human evaluation or at least an independent judge not used anywhere in the pipeline, together with a bias check (e.g., which tokens/features drive judge preferences).
  4. [Tables 2–3, Tables 5–6] Most of the reported gains over the strongest baseline PA-FAS are small in absolute terms: e.g., coarse ACC 98.75 vs 97.93, attack ACC 93.33 vs 91.82, and cross-domain ACC differences of roughly ±1 pp in some protocols. No error bars, multiple seeds, or significance tests are reported for any of the tables. Given the stochasticity of RL training and the small margins, this is insufficient to support a state-of-the-art claim. The authors should report mean ± std over at least three seeds, or otherwise justify why the comparisons are stable. This is especially important because the re-implemented baselines may also vary with seed.
minor comments (6)
  1. [Table 2, fine-grained ACC] The text says 'Bona fide samples are correct by construction, so this score should be read together with authenticity accuracy.' This makes the fine-grained attack-type accuracy difficult to interpret. Please state explicitly whether the denominator for fine-grained ACC includes bona fide samples and, if so, how the trivial 'real' class is handled.
  2. [Eq. (5) and Algorithm 1] Token indices are suppressed in the policy ratio r_{i,j}; please clarify whether the ratio is computed over full trajectories and whether the KL term is per-token. Also, Algorithm 1 refers to the degraded operator A but not to how n_c/n_d is chosen; please define these in the pseudocode.
  3. [Fig. 6(b)] The caption text '167 W → C / 1 C → W' is unclear/truncated. It should state the actual protocol and the meaning of 'hard' vs 'easy' subgroups.
  4. [AP@40/AP@50 definition] Please define how AP is computed for coarse boxes: whether it is the fraction of test samples with IoU > threshold (as appears to be the case) or a precision-recall area under the curve. The current notation is ambiguous.
  5. [Reproducibility] The paper says 'The code will be released soon' but no code, data, or trained checkpoints are available. FAS-R1-23K is a claimed contribution, so its release plan should be stated concretely (license, access, annotation format).
  6. [Table 4] The judge scores (e.g., Gemini-3-Pro judging its own outputs) may be inflated because the judge is also a candidate. This is related to Major Comment 3 and should at least be acknowledged in the text.

Circularity Check

1 steps flagged · score 4.0 of 10

One headline metric (attack-type ACC) has a definitional component; core authenticity and cross-domain results are grounded in external labels, so circularity is limited.

  1. self definitional [Experiments / In-Domain Evaluation, 'Fine-grained classification' paragraph]
    "Fine-grained classification. FAS-R1 reaches 93.33% ACC, slightly above FaceShield (Wang et al. 2025b). Bona fide samples are correct by construction, so this score should be read together with authenticity accuracy."

    For bona fide samples the attack-type output ('no attack') is correct whenever the authenticity decision is 'real'. Hence the 93.33% fine-grained attack-type ACC contains the 98.75% authenticity ACC as a guaranteed component for real samples, making the fine-grained metric partly a restatement of the coarse authenticity claim rather than an independent prediction of attack-type discrimination.

full rationale

The central derivation is not circular. Authenticity, attack-type, and localization labels are inherited from WMCA/PADISI-Face/SiW-Mv2 rather than generated by the model (Method, Data Construction: 'Authenticity labels, attack categories, and localization annotations are inherited from the original datasets rather than generated by MLLMs'), so the main ACC/AP numbers are measured against external ground truth. The two-stage SFT + GRPO pipeline (Eqs. 1-8) is a training procedure, not a fitted quantity renamed as a prediction; DSA and DA-GRPO are optimization components, and the ablations compare them on held-out metrics. The self-citation to FaceShield sets the evaluation protocol and a baseline but is not the load-bearing justification for FAS-R1's performance. Two evaluation concerns should be distinguished from circularity: (1) the in-domain protocol splits merged images at image level ('the merged images are split at the image level into training, validation, and test sets with an 8:1:1 ratio'), so identities overlap between train and test; this is an identity-leakage threat to generalization claims, not a definitional circularity because test labels are still external. (2) The answer-and-rationale quality evaluation is self-referential in character: training rationales were generated by Gemini 2.5 Flash, the RL reward includes 'GPT-5-based rationale-answer consistency', and the final quality scores are assigned by Gemini-3-Pro and Claude-Sonnet-4.5, which also appear as candidate systems ('Both judge models are also included as candidate systems and assign FAS-R1 higher scores than their own outputs'). This weakens the rationale-quality claim as independent evidence, but no equation forces the Elo outcome, so it is not counted as a formal circular step. The only construction-level dependency is the fine-grained attack-type metric, which the paper itself discloses as correct by construction for bona fide samples. Score 4 reflects this partial self-definitional component while recognizing that the headline authenticity and cross-domain results are externally grounded.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims depend on inherited label correctness, LLM-based dataset verification, the relevance of simulated degradations to real-world shifts, and unbiased LLM judging. These are domain assumptions rather than derived results, and none are independently verified with external evidence in the paper.

free parameters (3)
  • Reward weights for QA and localization = QA: 0.1/0.5/0.4; Loc: 0.2/0.8; tau=0.5
    Chosen by hand in the reward design; the RL optimization and all reported gains depend on these weights, and no sensitivity analysis is provided.
  • DA-GRPO hyperparameters = warm-up=5; rho=0.98; lambda=2.0; w_min=0.7; w_max=2.0; min prompts=4
    Chosen by hand and used to reweight advantages; the central claim that DA-GRPO improves hard subgroups depends on these values, with no ablation over them.
  • DSA degradation severity = moderate brightness/contrast/gamma/noise/JPEG
    The paper says 'moderate' but does not give quantitative ranges or sampling schedules; this is a free choice that directly affects the paired-rollout training signal and reproducibility.
assumptions (4)
  • domain assumption Original labels and attack-region boxes in WMCA, PADISI-Face, and SiW-Mv2 are correct and consistently annotated across datasets.
    These labels are inherited as ground truth for dataset construction, reward computation, and evaluation. If they contain noise or cross-dataset taxonomy inconsistencies, the reported metrics inherit that noise.
  • domain assumption GPT-5 verification and rule-based filtering remove hallucinated, inconsistent, or low-quality rationales.
    The dataset is claimed to be high-quality largely on the basis of this verification; the paper provides no quantitative verification of filter precision or recall, and manual inspection is applied only to flagged samples.
  • domain assumption The degradation operators used by DSA simulate real-world visual-quality shifts encountered in cross-domain deployment.
    DSA's central premise is that clean/degraded paired rollouts teach stable spoof cues; the paper does not validate that the chosen augmentations correspond to the actual distribution shift in the target domains.
  • domain assumption LLM judges (Claude-Sonnet-4.5 and Gemini-3-Pro) produce unbiased rationale-quality scores.
    The answer-and-rationale quality claim rests entirely on these judge models. Since they are also candidate systems and one judge family generated the training rationales, bias is plausible and no human evaluation is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FAS-R1: A Unified Multi-Task MLLM for Reasoning Face Anti-Spoofing." pith.science (2026). https://pith.science/paper/55LRRMEZ

@misc{pith2026260726432,
  author       = {Pith},
  title        = {Pith review of: FAS-R1: A Unified Multi-Task MLLM for Reasoning Face Anti-Spoofing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/55LRRMEZ}},
  note         = {Machine review of arXiv:2607.26432}
}
read the original abstract

Face anti-spoofing (FAS) is increasingly expected to provide not only bona fide/spoof decisions, but also attack semantics and image-grounded evidence for human inspection. Existing discriminative FAS models remain largely label-centric, while recent MLLM-based methods offer structured outputs but still rely mainly on supervised fine-tuning, often producing template-like rationales and weak optimization for difficult attacks. We propose FAS-R1, a two-stage reasoning-oriented MLLM framework for unified FAS prediction, covering authenticity classification, attack-type recognition and spoof-region localization. FAS-R1 first uses FAS-R1-23K, a high-quality long-CoT dataset, for cold-start supervised fine-tuning, and then performs FAS-specific GRPO post-training. Degradation-Simulated Augmentation (DSA) encourages stable spoof-cue reasoning across visual-quality shifts, while Difficulty-Aware GRPO (DA-GRPO) mitigates easy-sample dominance that may leave difficult task--attack groups under-optimized, especially for subtle or ambiguous attacks such as makeup and mask attacks. The main 3B FAS-R1 model achieves 98.75\% authenticity accuracy, 93.33\% attack-type accuracy, and 96.30/94.73\% AP@40/AP@50 in-domain. It also outperforms the compared systems in cross-domain authenticity generalization and answer-and-rationale quality. Experiments with different base models further show favorable scaling behavior. The code will be released soon.

Figures

Figures reproduced from arXiv: 2607.26432 by the authors.

Figure 1
Figure 1. Comparison of three FAS paradigms: (a) discrimi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Stage 1: long-CoT cold start with annotation-constrained generation, external verification, and rule/manual filtering. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Stage 2: FAS-specific reinforcement optimization. DSA constructs paired clean/degraded rollouts, while DA-GRPO [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: FAS-specific GRPO optimization. DSA mixes clean and degraded trajectories in each rollout group, and DA-GRPO [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Automated pairwise answer-and-rationale compar [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: FAS-specific RL diagnostics: (a) DSA improves [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Authenticity example on a replay attack. FAS-R1 identifies image-specific reflective cues and correctly predicts spoof, [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Localization example on a partial-eye attack. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

82 extracted references · 21 linked inside Pith

  1. [1]

    FirstName Alpher , title =

  2. [2]

    Journal of Foo , volume = 13, number = 1, pages =

    FirstName Alpher and FirstName Fotheringham-Smythe , title =. Journal of Foo , volume = 13, number = 1, pages =

  3. [3]

    Journal of Foo , volume = 14, number = 1, pages =

    FirstName Alpher and FirstName Fotheringham-Smythe and FirstName Gamow , title =. Journal of Foo , volume = 14, number = 1, pages =

  4. [4]

    FirstName Alpher and FirstName Gamow , title =

  5. [5]

    Computer Vision -- ECCV 2022 , year =

    2022

  6. [6]

    International Joint Conference on Biometrics (IJCB) , year =

    Face Spoofing Detection from Single Images Using Micro-Texture Analysis , author =. International Joint Conference on Biometrics (IJCB) , year =

  7. [7]

    BIOSIG , year =

    On the Effectiveness of Local Binary Patterns in Face Anti-Spoofing , author =. BIOSIG , year =

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Searching Central Difference Convolutional Networks for Face Anti-Spoofing , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

Show all 82 references
  1. [9]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Towards Unsupervised Domain Generalization for Face Anti-Spoofing , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  2. [10]

    Proceedings of the 38th International Conference on Machine Learning (ICML) , year =

    Learning Transferable Visual Models From Natural Language Supervision , author =. Proceedings of the 38th International Conference on Machine Learning (ICML) , year =

  3. [11]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    FLIP: Cross-domain Face Anti-spoofing with Language Guidance , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  4. [12]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    CFPL-FAS: Class Free Prompt Learning for Generalizable Face Anti-spoofing , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  5. [13]

    European Conference on Computer Vision (ECCV) , year =

    Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing , author =. European Conference on Computer Vision (ECCV) , year =

  6. [14]

    arXiv preprint arXiv:2311.18420 , year =

    TeG-DG: Textually Guided Domain Generalization for Face Anti-Spoofing , author =. arXiv preprint arXiv:2311.18420 , year =

    arXiv

  7. [15]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Visual Instruction Tuning , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  8. [16]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Training Language Models to Follow Instructions with Human Feedback , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  9. [17]

    arXiv preprint arXiv:2305.18290 , year =

    Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author =. arXiv preprint arXiv:2305.18290 , year =. 2305.18290 , archivePrefix =

    arXiv

  10. [18]

    arXiv preprint arXiv:1707.06347 , year =

    Proximal Policy Optimization Algorithms , author =. arXiv preprint arXiv:1707.06347 , year =. 1707.06347 , archivePrefix =

    arXiv

  11. [19]

    arXiv preprint arXiv:2402.03300 , year =

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author =. arXiv preprint arXiv:2402.03300 , year =. 2402.03300 , archivePrefix =

    arXiv

  12. [20]

    arXiv preprint arXiv:2501.12948 , year =

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning , author =. arXiv preprint arXiv:2501.12948 , year =. 2501.12948 , archivePrefix =

    arXiv

  13. [21]

    arXiv preprint arXiv:2503.06749 , year =

    Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models , author =. arXiv preprint arXiv:2503.06749 , year =. 2503.06749 , archivePrefix =

    arXiv

  14. [22]

    arXiv preprint arXiv:2503.12937 , year =

    R1-VL: Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization , author =. arXiv preprint arXiv:2503.12937 , year =. 2503.12937 , archivePrefix =

    arXiv

  15. [23]

    arXiv preprint arXiv:2504.18458 , year =

    Fast-Slow Thinking for Large Vision-Language Model Reasoning , author =. arXiv preprint arXiv:2504.18458 , year =. 2504.18458 , archivePrefix =

  16. [24]

    arXiv preprint arXiv:2504.10458 , year =

    GUI-R1: A Generalist R1-Style Vision-Language Action Model For GUI Agents , author =. arXiv preprint arXiv:2504.10458 , year =. 2504.10458 , archivePrefix =

    arXiv

  17. [25]

    arXiv preprint arXiv:2505.14460 , year =

    VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank , author =. arXiv preprint arXiv:2505.14460 , year =. 2505.14460 , archivePrefix =

  18. [26]

    arXiv preprint arXiv:2403.13372 , year =

    LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models , author =. arXiv preprint arXiv:2403.13372 , year =. 2403.13372 , archivePrefix =

    arXiv

  19. [27]

    arXiv preprint arXiv:2502.13923 , year =

    Qwen2.5-VL Technical Report , author =. arXiv preprint arXiv:2502.13923 , year =. 2502.13923 , archivePrefix =

    arXiv

  20. [28]

    arXiv preprint arXiv:2505.09415 , year =

    FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models , author =. arXiv preprint arXiv:2505.09415 , year =. 2505.09415 , archivePrefix =

  21. [29]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Deep Residual Learning for Image Recognition , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  22. [30]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    PatchNet: A Simple Face Anti-Spoofing Framework via Fine-Grained Patch Recognition , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  23. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Conditional Prompt Learning for Vision-Language Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  24. [32]

    arXiv preprint arXiv:2308.12966 , year =

    Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond , author =. arXiv preprint arXiv:2308.12966 , year =

    arXiv

  25. [33]

    arXiv preprint arXiv:2304.10592 , year =

    MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models , author =. arXiv preprint arXiv:2304.10592 , year =

    arXiv

  26. [34]

    arXiv preprint arXiv:2402.11530 , year =

    Efficient Multimodal Learning from Data-centric Perspective , author =. arXiv preprint arXiv:2402.11530 , year =

    arXiv

  27. [35]

    2025 , howpublished =

    Introducing Claude Sonnet 4.5 , author =. 2025 , howpublished =

    2025

  28. [36]

    2025 , howpublished =

    Update to GPT-5 System Card: GPT-5.2 , author =. 2025 , howpublished =

    2025

  29. [37]

    2025 , howpublished =

    Gemini models: Gemini 3 Pro , author =. 2025 , howpublished =

    2025

  30. [38]

    arXiv preprint arXiv:2312.02433 , year =

    Lenna: Language Enhanced Reasoning Detection Assistant , author =. arXiv preprint arXiv:2312.02433 , year =. 2312.02433 , archivePrefix =

    arXiv

  31. [39]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Instance-Aware Domain Generalization for Face Anti-Spoofing , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  32. [40]

    arXiv preprint arXiv:2409.03501 , year =

    Towards Data-Centric Face Anti-Spoofing: Improving Cross-domain Generalization via Physics-based Data Synthesis , author =. arXiv preprint arXiv:2409.03501 , year =

    arXiv

  33. [41]

    arXiv preprint arXiv:2311.07575 , year=

    Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models , author=. arXiv preprint arXiv:2311.07575 , year=

    arXiv

  34. [42]

    International Conference on Learning Representations (ICLR) , year =

    Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning , author =. International Conference on Learning Representations (ICLR) , year =

  35. [43]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Interpretable face anti-spoofing: Enhancing generalization with multimodal large language models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  36. [44]

    2026 , eprint =

    Harnessing Chain-of-Thought Reasoning in Multimodal Large Language Models for Face Anti-Spoofing , author =. 2026 , eprint =

    2026

  37. [45]

    2025 , eprint =

    Exploring Task-Solving Paradigm for Generalized Cross-Domain Face Anti-Spoofing via Reinforcement Fine-Tuning , author =. 2025 , eprint =

    2025

  38. [46]

    2025 , eprint =

    RAIDX: A Retrieval-Augmented Generation and GRPO Reinforcement Learning Framework for Explainable Deepfake Detection , author =. 2025 , eprint =

    2025

  39. [47]

    2025 , eprint =

    VidGuard-R1: AI-Generated Video Detection and Explanation via Reasoning MLLMs and RL , author =. 2025 , eprint =

    2025

  40. [48]

    2026 , eprint =

    Interpretable All-Type Audio Deepfake Detection with Audio LLMs via Frequency Time-Group Relative Policy Optimization , author =. 2026 , eprint =

    2026

  41. [49]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Fast-Slow Thinking GRPO for Large Vision-Language Model Reasoning , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  42. [50]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  43. [51]

    International Conference on Learning Representations (ICLR) , year =

    DIVA-GRPO: Enhancing Multimodal Reasoning through Difficulty-Adaptive Variant Advantage , author =. International Conference on Learning Representations (ICLR) , year =

  44. [52]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Towards Unified Multimodal Interleaved Generation via Group Relative Policy Optimization , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  45. [53]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Q-Insight: Understanding Image Quality via Visual Reinforcement Learning , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  46. [54]

    2025 , eprint =

    Seeing is Believing? Mitigating OCR Hallucinations in Multimodal Large Language Models , author =. 2025 , eprint =

    2025

  47. [55]

    2025 , eprint =

    REVEAL: Reasoning-enhanced Forensic Evidence Analysis for Explainable AI-generated Image Detection , author =. 2025 , eprint =

    2025

  48. [56]

    2025 , eprint =

    DeepVideo-R1: Video Reinforcement Fine-Tuning via Difficulty-aware Regressive GRPO , author =. 2025 , eprint =

    2025

  49. [57]

    2025 , eprint =

    MedVLM-R1: Incentivizing Medical Reasoning Capability of Vision-Language Models (VLMs) via Reinforcement Learning , author =. 2025 , eprint =

    2025

  50. [58]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Deep learning for face anti-spoofing: A survey , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2022 , publisher=

    2022

  51. [59]

    Computers, Materials & Continua , year =

    PNSS: Unknown Face Presentation Attack Detection with Pseudo Negative Sample Synthesis , author =. Computers, Materials & Continua , year =

  52. [60]

    IEEE International Conference on Image Processing (ICIP) , pages =

    Face Anti-Spoofing Based on Color Texture Analysis , author =. IEEE International Conference on Image Processing (ICIP) , pages =

  53. [61]

    International Joint Conference on Biometrics (IJCB) , pages =

    Face Anti-Spoofing Using Patch and Depth-Based CNNs , author =. International Joint Conference on Biometrics (IJCB) , pages =. 2017 , doi =

    2017

  54. [62]

    International Conference on Biometrics (ICB) , year =

    Deep Pixel-wise Binary Supervision for Face Presentation Attack Detection , author =. International Conference on Biometrics (ICB) , year =

  55. [63]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Face Anti-Spoofing: Model Matters, so Does Data , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2019 , doi =

    2019

  56. [64]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Deep Tree Learning for Zero-Shot Face Anti-Spoofing , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  57. [65]

    IEEE Transactions on Information Forensics and Security , volume =

    Biometric Face Presentation Attack Detection with Multi-Channel Convolutional Neural Network , author =. IEEE Transactions on Information Forensics and Security , volume =

  58. [66]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Detection and Continual Learning of Novel Face Presentation Attacks , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  59. [67]

    arXiv preprint arXiv:2208.11148 , year =

    Multi-domain Learning for Updating Face Anti-spoofing Models , author =. arXiv preprint arXiv:2208.11148 , year =

    arXiv

  60. [68]

    2025 , howpublished =

    Gemini 2.5 Flash , author =. 2025 , howpublished =

    2025

  61. [69]

    2025 , howpublished =

    GPT-5 , author =. 2025 , howpublished =

    2025

  62. [70]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , year =

    GRPO-LEAD: A Difficulty-Aware Reinforcement Learning Approach for Concise Mathematical Reasoning in Language Models , author =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , year =

    2025

  63. [71]

    Visual Intelligence , year =

    SHIELD: An Evaluation Benchmark for Face Spoofing and Forgery Detection with Multimodal Large Language Models , author =. Visual Intelligence , year =. 2402.04178 , archivePrefix =

    arXiv

  64. [72]

    2026 , eprint =

    From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing , author =. 2026 , eprint =

    2026

  65. [73]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    FSFM: A Generalizable Face Security Foundation Model via Self-Supervised Facial Representation Learning , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  66. [74]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Gradient Alignment for Cross-Domain Face Anti-Spoofing , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =. 2402.18817 , archivePrefix =

    arXiv

  67. [75]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    PA-FAS: Towards Interpretable and Generalizable Multimodal Face Anti-Spoofing via Path-Augmented Reinforcement Learning , author =. Proceedings of the AAAI Conference on Artificial Intelligence , year =. 2511.17927 , archivePrefix =

  68. [76]

    2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Domain randomization for transferring deep neural networks from simulation to the real world , author=. 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2017 , organization=

    2017

  69. [77]

    arXiv preprint arXiv:2004.14990 , year=

    Reinforcement Learning with Augmented Data , author=. arXiv preprint arXiv:2004.14990 , year=

    2004 arXiv

  70. [78]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Multi-View Slot Attention Using Paraphrased Texts for Face Anti-Spoofing , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  71. [79]

    International Conference on Learning Representations , year =

    Harder Is Better: Boosting Mathematical Reasoning via Difficulty-Aware GRPO and Multi-Aspect Question Reformulation , author =. International Conference on Learning Representations , year =

  72. [80]

    arXiv preprint arXiv:2602.21743 , year =

    Enhancing Multi-Modal LLMs Reasoning via Difficulty-Aware Group Normalization , author =. arXiv preprint arXiv:2602.21743 , year =

  73. [81]

    Yu, Qiying and Zhang, Zheng and Zhu, Ruofei and Yuan, Yufeng and Zuo, Xiaochen and Yue, Yu and Dai, Weinan and Fan, Tiantian and Liu, Gaohong and Liu, Lingjun and Liu, Xin and Lin, Haibin and Lin, Zhiqi and Ma, Bole and Sheng, Guangming and Tong, Yuxuan and Zhang, Chi and Zhan...

  74. [82]

    arXiv preprint arXiv:2507.18071 , year =

    Group Sequence Policy Optimization , author =. arXiv preprint arXiv:2507.18071 , year =

    arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.