Pith. sign in

REVIEW 3 major objections 7 minor 94 references

AI-image detectors generalize better when they first learn three checkable perception skills, then fold them into reasoning through value-aware distillation.

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

2026-07-30 10:54 UTC pith:3GJQ3V4M

load-bearing objection Solid recipe paper: cleaned cold-start + verifiable perception RL + value-aware OPD actually moves the needle on MLLM AIGI detection; the causal “perception is the foundation” story is softer than the framing. the 3 major comments →

arxiv 2607.27113 v1 pith:3GJQ3V4M submitted 2026-07-29 cs.CV

Veritas++: Value-aware On-Policy Distillation for Perception-Enhanced AIGI Detection

Hao Tan , Jun Lan , Zichang Tan , Ajian Liu , Zijian Yu , Chuanbiao Song , Huijia Zhu , Weiqiang Wang
show 2 more authors
Jun Wan Zhen Lei
This is my paper
classification cs.CV
keywords AI-generated image detectionmultimodal large language modelsperceptionon-policy distillationfine-grained anomaliessemantic anomaliesgeneralizationverifiable rewards
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.

Multimodal language models used as AI-generated image detectors still miss fine-grained visual anomalies, even after detection fine-tuning, because training has focused on how explanations are written rather than on seeing. This paper argues that authenticity judgment should rest on three basic perception abilities: noticing small visual details, spotting semantic anomalies such as distorted anatomy, and detecting pixel-level discrepancies. It strengthens those abilities with Perception-oriented Learning that replaces open-ended description labels with verifiable accuracy and localization rewards, then internalizes the improved perception into pattern-aware reasoning via Value-aware On-Policy Distillation, which upweights high-value signals on the model’s own rollouts. Across standard, in-the-wild, and emerging-generator benchmarks the resulting system reports strong accuracy, and a small targeted update can add new forgeries without large losses on prior settings. A sympathetic reader would care because generators and real-world contexts keep changing, so detectors need transferable perception rather than brittle artifact scripts.

Core claim

Current MLLM-based AI-generated image detectors are limited by weak intrinsic perception of fine-grained details, semantic anomalies, and pixel discrepancies. Explicitly training those three capacities with verifiable rewards bridges that gap and yields seamless detection gains; Value-aware On-Policy Distillation then internalizes the enhanced perception into authenticity reasoning and supports efficient adaptation to novel generators without sacrificing existing performance.

What carries the argument

Perception-oriented Learning (PoRL) and Value-aware On-Policy Distillation (VaOPD). PoRL trains fine-grained, semantic-anomaly, and pixel-discrepancy tasks under checkable accuracy and bounding-box rewards; VaOPD uses a privileged self-teacher on student rollouts and reweights distillation at trajectory, token, and direction levels so corrective, high-value signals dominate.

Load-bearing premise

That gains on the three proxy perception tasks with checkable rewards are what actually cause better authenticity reasoning, rather than side effects of extra training, cleaner cold-start data, or privileged labels in the teacher.

What would settle it

Run the same three-stage pipeline with perception rewards scrambled or removed while holding data volume, compute, cold-start filtering, and distillation structure fixed; if standard, in-the-wild, and emerging detection accuracy stay essentially unchanged, the claim that those perception skills are the foundation of the gains is falsified.

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

If this is right

  • Detectors that only supervise explanation text without verifiable perception will keep missing subtle and novel forgeries.
  • Small, value-weighted distillation updates can add capability for new generators and scenarios without large forgetting on prior benchmarks.
  • Fine-grained, semantic, and pixel perception are complementary; removing any branch hurts some evaluation regime.
  • Reasoning traces become more concrete and less globally plausible when grounded in trained perception of local and low-level evidence.
  • Transparent MLLM detectors can match or exceed strong vision-only systems on balanced OOD and in-the-wild suites once perception is trained first.

Where Pith is reading between the lines

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

  • The same three-skill perception curriculum could transfer to finer forensic outputs such as anomaly localization and generator attribution without redesigning the full reasoning scaffold.
  • Value-aware on-policy distillation is a plausible general recipe for continual multimodal adaptation whenever new domains arrive faster than full retraining.
  • Failure cases on semantic illusions and heavily filtered real photos imply that broadening authentic-image diversity may matter as much as covering new synthetic generators.
  • If real generator artifacts drift away from the current proxy tasks, the verifiable task suite itself would need periodic refresh rather than one-shot perception training.

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. The manuscript presents Veritas++, an MLLM-based AI-generated image (AIGI) detector built on Qwen3-VL-8B with a three-stage pipeline: (1) a human-annotation-only cold start rewritten into pattern-aware reasoning chains; (2) Perception-oriented Learning (PoRL), an RLVR stage (GSPO) on three proxy perception tasks with verifiable rewards — fine-grained region QA (ZwZ/Vero), semantic anomaly judgment (AbHuman, MagicData), and pixel-discrepancy localization (Next-IMDL pairs); and (3) Value-aware On-Policy Distillation (VaOPD), in which a privileged EMA self-teacher (conditioned on labels/guidance) provides dense token-level supervision on student rollouts, with trajectory reweighting toward incorrect rollouts, token reweighting by teacher–student KL, and an adaptive JSD direction driven by an entropy-gap/advantage corrective signal. Across standard, in-the-wild, and emerging benchmarks the authors report ~94%/~90%/~70–81% average accuracy, improving over their prior Veritas by ~15–17 points and over reproduced vision-only and MLLM baselines, and show that adding small amounts of TFR/GPT-Image-2 data at the VaOPD stage yields large emerging-regime gains with minimal forgetting. Ablations (Tables IV–VI) decompose the pipeline, the perception tasks, and the VaOPD components.

Significance. If the results hold, this is a solid empirical contribution to MLLM-based forensic detection: the evaluation is unusually broad (10+ benchmarks across three difficulty regimes, including GPT-Image-2 and text-rich forgeries), baselines are reproduced from official checkpoints rather than copied numbers, and the staged ablations in Tables IV–VI plus the VaOPD-vs-OPD-vs-GSPO comparison in Table VI are genuinely informative — the GSPO row demonstrating catastrophic forgetting under RLVR is a useful result in its own right. The VaOPD mechanism (value-weighted privileged self-distillation with adaptive JSD direction) is a reasonable, clearly specified contribution to the on-policy distillation literature, and the promised release of code and checkpoints supports reproducibility. The weaker link is the causal narrative that perception learning is the *foundation* of the detection gains: the evidence for that specific claim is currently partly in-distribution and partly confounded (see major comments), though the end-task detection improvements stand independently of how that narrative resolves.

major comments (3)
  1. [Table I / §III-B] The strongest quantitative evidence for 'bridging the perception gap' is partly in-distribution. Table I reports perception gains of 53.1→96.6 average, but §III-B states that PoRL trains directly on AbHuman (region-level anomaly) and MagicData (image-level anomaly) — precisely the two columns with the most dramatic jumps (97.4 and 99.4). Near-ceiling scores on training distributions are consistent with task fitting rather than a generalizable perception capability. Only SDG and B-Free are held out, and the fine-grained branch (ZwZ/Vero) is not evaluated in Table I at all. Since the paper's central framing (abstract, §I, §IV-C) is that enhanced perception causes the detection gains, the authors should add a genuinely held-out perception evaluation: e.g., evaluate all three perception skills on datasets disjoint from the PoRL training sources (including a held-out fine-grained and a held-o
  2. [Table IV / §IV-C] The attribution of detection gains to perception learning is confounded by the absence of a compute-matched RL control. The '+SFT+PoRL' vs '+SFT' comparison (+2.3% standard, +3.3% emerging Acc) attributes the gain to perception, but PoRL is an entire additional RL stage (20K samples, GSPO, G=4 rollouts). RLVR on a cold-started MLLM with verifiable rewards on perception-irrelevant data (e.g., math or generic VQA) could plausibly produce part of this gain via improved calibration or answer formatting. Note also that on the emerging regime, PoRL raises Acc but *lowers* F1 (54.7→53.6), which the 'seamless improvements' framing does not acknowledge. A control row '+SFT+RLVR(perception-irrelevant)' with matched compute would isolate the causal contribution of the perception tasks; without it, the 'perception as foundation' claim should be tempered. This is fixable within the paper's existing i
  3. [§IV-A / Tables II–III († rows) / Stage-3 data] The training-data provenance for Stage 3 (VaOPD, 11K samples) is not specified, and the † 'capability evolution' evaluation changes the meaning of 'emerging OOD'. For VERITAS++†, TFR and GPT-Image-2 training samples are incorporated, after which TFR (+9.4) and GPT-Image-2 (+11.2) are no longer out-of-distribution — the 80.6 'Emerging OOD Avg' for the † model is therefore not comparable to the same column for other methods and should be relabeled or reported separately. More importantly, the manuscript never states which detection data the base VaOPD stage uses; given that Stage 1 draws on HydraFake and Table II evaluates on HydraFake's cross-forgery set, the authors should explicitly list all Stage-3 training sources and confirm that no evaluation benchmark's test split (or near-duplicates) appears in any stage's training data. This is a leakage-control clarification that the current tex
minor comments (7)
  1. [Eq. (5)] The area penalty is discontinuous at the boundary: for ρ(ˆb,b)≤γ the penalty is 0, but immediately above γ it jumps to −δ[log₂(ρ/γ)+1] = −δ = −0.3. A smooth ramp (e.g., −δ·log₂(ρ/γ)) would avoid a cliff that the policy could exploit or oscillate around; at minimum, please justify the design.
  2. [Table I] The 'Detection' column header does not state which benchmark or metric it reports. Please specify (dataset, Acc/F1) in the caption.
  3. [§III-B, Eq. (2)] The LLM-as-a-Judge fallback for the fine-grained reward is underspecified: which judge model, what fraction of rewards are adjudicated by it, and what agreement it has with deterministic matching? Judge identity matters because reward noise here directly shapes the perception policy.
  4. [§IV-A] No sensitivity analysis is provided for the new hyperparameters (τ_r=0.5, [β_min,β_max]=[0.1,0.5], η=0.99, γ=2.0, δ=0.3, L_max/L_cache). Even a small sweep of τ_r and the β range on one benchmark would strengthen Table VI's component claims.
  5. [Table VII / §IV-D] The reasoning-quality evaluation uses 100 samples and a single MLLM judge (Gemini3.0-Pro-Preview); Elo differences of this size on n=100 are noisy. Please report confidence intervals or a second judge, and note that BusterX++'s higher absolute Physics score (4.13 vs 4.08) complicates the claim of uniform improvement.
  6. [Figure 4] The shaded 'one standard deviation' bands do not state over what population the variance is computed (tokens within a trajectory? across trajectories? across training steps?). Please clarify, and state how many rollouts were probed.
  7. [Writing] Typos/grammar: 'hiqh-quality' (§III opening), 'Combing enhanced visual perception' (contributions), 'raising persistent concerns about and digital deception' (§I), 'the query asks the model to determine the enclosed body part is normal' (§III-B). A proofreading pass is recommended.

Circularity Check

1 steps flagged

Empirical methods paper with external held-out detection benchmarks; no derivation-by-construction circularity.

specific steps
  1. fitted input called prediction [Table I; §III-B Semantic anomaly perception; Abstract claim on bridging perception gap]
    "VERITAS++(Percept.) 97.4 97.5 99.4 92.0 96.6 86.9 ... (1) For region-level judgment, the training data is sampled from AbHuman [21]. ... (2) For image-level recognition, the training data is sampled from MagicData [22]. ... The perception learning effectively bridges the perception gap and yields seamless gains on detection"

    Two of the four perception columns used to quantify the 'bridged perception gap' (AbHuman, MagicData) are the same sources PoRL trains on, so near-ceiling scores there are partly training-task fit. This mildly inflates the perception-gap narrative; it does not force the external detection numbers, which are measured on held-out AIGI benchmarks.

full rationale

Veritas++ is a three-stage empirical ML pipeline (cold-start SFT, PoRL with verifiable rewards, VaOPD). Its load-bearing claims are accuracy/F1 numbers on external standard, in-the-wild, and emerging AIGI benchmarks (Chameleon, GenBuster++, HydraFake, WildRF, TFR, GPT-Image-2, etc.), not closed-form identities. Self-citation of VERITAS is used as a prior baseline and architectural starting point, not as a uniqueness theorem that forces the present results. The privileged self-teacher (training-only labels/guidance) is standard on-policy distillation asymmetry, not a circular proof. Table I does report near-ceiling perception scores on AbHuman and MagicData, which §III-B also uses as PoRL training sources—so those two columns are partly in-distribution task fit rather than pure transfer—but that is weak evidence presentation, not a prediction forced by construction of the main detection claims, which remain externally measured (Tables II–VI). No self-definitional loop, no uniqueness import, no ansatz smuggled as theorem, and no renaming of a known law. Score 1 reflects only that mild in-distribution perception reporting, not central circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

Load-bearing content is methodological and empirical, not a formal derivation. The claim rests on domain assumptions about what perception proxies matter for AIGI detection, on RL/distillation design choices and hyperparameters, and on the validity of human/MLLM-curated supervision. No new physical entities; invented constructs are algorithmic modules (PoRL task suite, VaOPD weighting rules, privileged self-teacher).

free parameters (6)
  • VaOPD β range [β_min, β_max] = [0.1, 0.5]
    Controls adaptive distillation direction per token; set to [0.1, 0.5] by authors, directly shapes the objective.
  • Trajectory reweighting temperature τ_r = 0.5
    Softmax temperature for prioritizing incorrect rollouts; hand-chosen, changes which trajectories dominate the loss.
  • EMA decay η for self-teacher = 0.99
    Defines privileged teacher as exponential moving average of student; stability/quality of distillation depends on this choice.
  • Pixel localization area-penalty (γ, δ) = γ=2.0, δ=0.3
    IoU reward penalty hyperparameters to stop oversized boxes; ad hoc to the pixel perception task.
  • Length reward window (L_max, L_cache) and λ_len = L_max=1536, L_cache=896
    Soft length shaping in PoRL; affects verbosity and effective reward scale across perception tasks.
  • LoRA rank and α, stage sample counts/epochs = rank=64, α=128; ~10K×2 / 20K×1 / 11K×1
    Capacity and data budget knobs that condition all reported accuracies; not derived from theory.
axioms (6)
  • domain assumption Fine-grained VQA accuracy, semantic anomaly label accuracy, and pixel-edit localization IoU are valid, transferable proxies for the perceptual skills needed in general AIGI authenticity judgment.
    Core premise of PoRL (Stage 2); detection gains are interpreted as evidence that these proxies close the perception bottleneck.
  • domain assumption Human-annotated artifact evidence rewritten by Gemini into pattern-aware chains (then GPT-filtered) yields low-hallucination supervision without systematically distorting forensic cues.
    Stage 1 cold-start construction; quality of all later reasoning depends on this rewrite/filter pipeline.
  • domain assumption A privileged self-teacher conditioned on training-only labels and perceptual guidance provides denser correct perception-aware targets on student rollouts without collapsing to mere label memorization.
    VaOPD setup (Eqs. 12–13); required for the internalization and capability-evolution narrative.
  • standard math Group Sequence Policy Optimization (GSPO) with normalized group advantages is an appropriate RL objective for multi-task verifiable perception rewards.
    Uses standard on-policy RLVR machinery (Eqs. 10–11); not proved optimal here but taken from prior RL literature.
  • domain assumption Held-out public AIGI benchmarks and the authors’ GPT-Image-2 collection are sufficiently uncontaminated relative to training mixtures to support generalization claims.
    Implicit in all OOD tables; not independently audited in the paper.
  • ad hoc to paper Token-level teacher–student KL, entropy gap, and advantage identify ‘high-value’ distillation positions in a way that improves learning rather than overfitting teacher quirks.
    VaOPD’s trajectory/token/direction design (Eqs. 14–20) is paper-specific operationalization of ‘distillation value’.
invented entities (3)
  • Perception-oriented Learning (PoRL) task suite independent evidence
    purpose: Replace open-ended artifact description with three verifiable perception reward channels for AIGI-oriented MLLM training.
    Algorithmic training construct; evaluated via ablations, not an external physical entity.
  • Value-aware On-Policy Distillation (VaOPD) independent evidence
    purpose: Reweight trajectories/tokens and adapt JSD direction so privileged self-distillation internalizes perception-aware reasoning and supports capability evolution.
    New distillation mechanism defined by the paper’s weighting and β schedule; evidence is internal ablations vs OPD/GSPO.
  • Privileged self-teacher (EMA + labels/guidance) no independent evidence
    purpose: Provide dense token distributions on student rollouts without a separate external teacher model generating full answers.
    Training-time construct combining OPSD ideas with AIGI-specific privileged inputs.

pith-pipeline@v1.2.0-grok45-kimik3 · 31866 in / 4418 out tokens · 89444 ms · 2026-07-30T10:54:29.730381+00:00 · methodology

0 comments
read the original abstract

The growing capability of image generation models has made synthetic images a routine presence in open media, making robust and generalizable AI-Generated Image (AIGI) detection increasingly essential. While multi-modal large language models (MLLMs) offer a transparent alternative to black-box binary scoring, we observe that current MLLM-based detectors still exhibit notable perception bottlenecks in capturing fine-grained anomalies. They primarily focus on how visual evidence is organized and synthesized, leaving the intrinsic perception less optimized. To mitigate this gap, we present Veritas++, a perception-enhanced reasoning framework that establishes reliable perception as the foundation of authenticity reasoning. Rather than directly optimizing the model's explanatory ability, we ground AIGI detection on three basic perception abilities, i.e., capturing fine-grained visual details, semantic anomalies and pixel-level differences. Building on this insight, we introduce Perception-oriented Learning (PoRL), which replaces open-ended description supervision with verifiable rewards to explicitly strengthen these capacities. To further integrate enhanced perception with reasoning, we introduce Value-aware On-Policy Distillation (VaOPD), an adaptive distillation mechanism that prioritizes high-value distillation signals over uniform supervision, internalizing perception-aware reasoning through a privileged self-teacher. Extensive experiments across standard, in-the-wild and emerging benchmarks demonstrate that Veritas++ achieves promising generalization. The perception learning effectively bridges the perception gap and yields seamless gains on detection, while VaOPD further enables efficient capability evolvement without sacrificing existing performance. Code and checkpoints are available at https://github.com/EricTan7/VeritasPP.

Figures

Figures reproduced from arXiv: 2607.27113 by Ajian Liu, Chuanbiao Song, Hao Tan, Huijia Zhu, Jun Lan, Jun Wan, Weiqiang Wang, Zhen Lei, Zichang Tan, Zijian Yu.

Figure 1
Figure 1. Figure 1: Paradigms of MLLM-based AIGI detection. (a) Explanation-supervised methods learn from artifact rationales. (b) Reasoning-oriented methods, including VERITAS [3], organize visual evidence into structured reasoning. (c) Evidence￾augmented methods incorporate external experts, tools, or cues. (d) VERITAS++explicitly strengthens fine-grained, semantic, and pixel-level perception, and internalizes it into reaso… view at source ↗
Figure 2
Figure 2. Figure 2: Examples of perception challenges in AIGI detec￾tion. (a) Fine-grained details and semantic anomalies require localized structural understanding. (b) Pixel-level discrepan￾cies can remain subtle even between paired images. VL are almost incapable of recognizing visual anomalies and pixel-level differences. (3) Detection fine-tuning alone could not effectively mitigate this gap. Building upon this insight, … view at source ↗
Figure 3
Figure 3. Figure 3: Overview of VERITAS++. Stage 1: High-quality Cold Start establishes a reliable, low-hallucination reasoning prior. We retain only images with manually annotated evidence and rewrite their short explanations or forgery masks into pattern￾aware reasoning chains. Stage 2: Learn to Perceive bridges the perception gap through targeted challenges on fine-grained visual objects, semantic anomalies and pixel-level… view at source ↗
Figure 4
Figure 4. Figure 4: Empirical motivation for VaOPD. Token-wise teacher and student entropy (top) and teacher advantage vt (bottom) vary substantially along the reasoning process. Shaded regions denote one standard deviation. Corrective signals concentrate at positions with higher teacher entropy and negative vt, whereas late tokens are often low-value. and reasoning alternatives not covered by the student. At such informative… view at source ↗
Figure 5
Figure 5. Figure 5: Overview of Value-aware On-Policy Distillation (VaOPD). Left: The student samples multiple on-policy reasoning trajectories under the standard authenticity query. Middle: An EMA self-teacher, additionally conditioned on training-only labels and perceptual guidance, scores the same student-generated token prefixes and provides dense privileged supervision without generating separate responses. Right: VaOPD … view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of GSPO, vanilla OPD and VaOPD. Standard and in-the-wild results are averaged over their cor￾responding benchmarks. VaOPD better balances existing gen￾eralization and capability acquisition on emerging scenarios. ing erroneous rollouts and informative tokens jointly focuses the supervision on the most learning-demanding positions. (3) The full combination of all three components achieves further… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison between VERITAS and VERITAS++. While VERITAS is misled by global plausibility, VERITAS++ identifies subtle distortions in the motorcycle and distant building (top), as well as irregular text and diagram details in the generated poster (bottom). Comparison of token distributions between VaOPD and OPD [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Representative failure cases of VERITAS++. A semantic illusion in a real photograph leads to a false positive (top), while a high-quality generated portrait embedded in a realistic live-stream context causes a false negative (bottom). the polished facial appearance is misattributed to common beautification filters, illustrating a typical false negative. These cases highlight the remaining difficulty in dis… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

94 extracted references · 31 linked inside Pith

  1. [1]

    Aigi-holmes: Towards explainable and generalizable ai-generated image detection via multimodal large language models,

    Z. Zhou, Y . Luo, Y . Wu, K. Sun, J. Ji, K. Yan, S. Ding, X. Sun, Y . Wu, and R. Ji, “Aigi-holmes: Towards explainable and generalizable ai-generated image detection via multimodal large language models,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 18 746–18 758

    2025

  2. [2]

    Spot the fake: Large multimodal model-based synthetic image detection with artifact explanation,

    S. Wen, P. Feng, H. Kang, Z. Wen, Y . Chen, J. Wu, C. He, W. Liet al., “Spot the fake: Large multimodal model-based synthetic image detection with artifact explanation,”Advances in Neural Information Processing Systems, vol. 38, pp. 58 972–59 005, 2026

    2026

  3. [3]

    Veritas: Generalizable deepfake detection via pattern-aware reasoning,

    H. Tan, J. Lan, Z. Tan, A. Liu, C. Song, S. Shi, H. Zhu, W. Wang, J. Wan, and Z. Lei, “Veritas: Generalizable deepfake detection via pattern-aware reasoning,” inInternational Conference on Learning Representations, 2026

    2026

  4. [4]

    Towards universal fake image detec- tors that generalize across generative models,

    U. Ojha, Y . Li, and Y . J. Lee, “Towards universal fake image detec- tors that generalize across generative models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 24 480–24 489

    2023

  5. [5]

    Orthogonal subspace decomposition for generalizable ai-generated image detection,

    Z. Yan, J. Wang, P. Jin, K.-Y . Zhang, C. Liu, S. Chen, T. Yao, S. Ding, B. Wu, and L. Yuan, “Orthogonal subspace decomposition for generalizable ai-generated image detection,” inInternational Conference on Machine Learning. PMLR, 2025, pp. 70 268–70 288

    2025

  6. [6]

    Dual data alignment makes AI-generated image detector easier generalizable,

    R. Chen, J. Xi, Z. Yan, K.-Y . Zhang, S. Wu, J. Xie, X. Chen, L. Xu, I. Guan, T. Yao, and S. Ding, “Dual data alignment makes AI-generated image detector easier generalizable,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  7. [7]

    Fakexplain: Ai-generated image detection via human-aligned grounded reasoning,

    Y . Ji, Y . Hong, Q. Fan, H. Zhu, W. Wang, L. Zhang, J. Zhanget al., “Fakexplain: Ai-generated image detection via human-aligned grounded reasoning,” inThe Fourteenth International Conference on Learning Representations, 2026

    2026

  8. [8]

    Fakeshield: Explainable image forgery detection and localization via multi-modal large language models,

    Z. Xu, X. Zhang, R. Li, Z. Tang, Q. Huang, and J. Zhang, “Fakeshield: Explainable image forgery detection and localization via multi-modal large language models,” inInternational Conference on Learning Rep- resentations, vol. 2025, 2025, pp. 31 186–31 216

    2025

  9. [9]

    Sida: Social media image deepfake detection, localization and explanation with large multimodal model,

    Z. Huang, J. Hu, X. Li, Y . He, X. Zhao, B. Peng, B. Wu, X. Huang, and G. Cheng, “Sida: Social media image deepfake detection, localization and explanation with large multimodal model,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 28 831– 28 841

    2025

  10. [10]

    Legion: Learning to ground and explain for synthetic image detection,

    H. Kang, S. Wen, Z. Wen, J. Ye, W. Li, P. Feng, B. Zhou, B. Wang, D. Lin, L. Zhanget al., “Legion: Learning to ground and explain for synthetic image detection,” inProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 2025, pp. 18 937–18 947

    2025

  11. [11]

    Fake-hr1: Rethinking reasoning of vision language model for synthetic image detection,

    C. Jiang, X. Sha, F. Yu, J. Liu, J. Liu, M. Fang, C. Zhang, and W. Lu, “Fake-hr1: Rethinking reasoning of vision language model for synthetic image detection,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 10 482–10 486

    2026

  12. [12]

    Fakevlm-r1: Internalizing physical laws via cot for synthetic image detection,

    L. Zhu, J. Ye, K. Lin, Z. Yan, C. He, and W. Li, “Fakevlm-r1: Internalizing physical laws via cot for synthetic image detection,”arXiv preprint arXiv:2605.30062, 2026

    Pith/arXiv arXiv 2026

  13. [13]

    Busterx++: Towards unified cross-modal ai-generated content detection and explanation with mllm,

    H. Wen, T. Li, Z. Huang, Y . He, and G. Cheng, “Busterx++: Towards unified cross-modal ai-generated content detection and explanation with mllm,”arXiv preprint arXiv:2507.14632, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Vigil: Part-grounded structured reasoning for generalizable deepfake detection,

    X. Li, J. Xu, and J. Chen, “Vigil: Part-grounded structured reasoning for generalizable deepfake detection,”arXiv preprint arXiv:2603.21526, 2026

    Pith/arXiv arXiv 2026

  15. [15]

    Realign: Gener- alizable image forgery detection via reasoning-aligned representation,

    Q. Huang, Z. Xu, X. Zhang, X. Yu, and J. Zhang, “Realign: Gener- alizable image forgery detection via reasoning-aligned representation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 21 295–21 305

    2026

  16. [16]

    Reveal: Reasoning-enhanced forensic evidence analysis for explainable ai-generated image detection,

    H. Cao, Q. Mei, Z. Li, Y . Li, Z. Meng, Y . Zhang, C. Li, Z. Zhang, X. Ding, Y . Wanget al., “Reveal: Reasoning-enhanced forensic evidence analysis for explainable ai-generated image detection,”arXiv preprint arXiv:2511.23158, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Evoguard: An extensible agentic rl-based framework for practical and evolving ai- generated image detection,

    C. Zhu, M. Wang, J. Liu, C.-C. Chang, and I. Echizen, “Evoguard: An extensible agentic rl-based framework for practical and evolving ai- generated image detection,”arXiv preprint arXiv:2603.17343, 2026

    Pith/arXiv arXiv 2026

  18. [18]

    Agentfox: Llm agent-guided fusion with explainability for ai-generated image detection,

    Y . Yu, Y . Zhou, B. Li, K. Lin, H. Li, J. Ni, and B. Cao, “Agentfox: Llm agent-guided fusion with explainability for ai-generated image detection,”arXiv preprint arXiv:2603.23115, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    Locate-then-examine: Grounded region reasoning improves detection of ai-generated images,

    Y . Ji, Y . Hong, B. Deng, J. Lan, H. Zhu, W. Wang, L. Zhang, and J. Zhang, “Locate-then-examine: Grounded region reasoning improves detection of ai-generated images,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 19 165–19 175

    2026

  20. [20]

    Unigendet: A unified generative-discriminative framework for co-evolutionary image generation and generated image detection,

    Y . Zhang, W. Zheng, Y . Li, B. Yu, Y . Zheng, L. Chen, J. Lu, and J. Zhou, “Unigendet: A unified generative-discriminative framework for co-evolutionary image generation and generated image detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 16 226–16 236

    2026

  21. [21]

    Humanrefiner: Benchmarking abnormal human generation and refining with coarse-to-fine pose-reversible guidance,

    G. Fang, W. Yan, Y . Guo, J. Han, Z. Jiang, H. Xu, S. Liao, and X. Liang, “Humanrefiner: Benchmarking abnormal human generation and refining with coarse-to-fine pose-reversible guidance,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 201–217

    2024

  22. [22]

    Magicmirror: A large-scale dataset and benchmark for fine-grained artifacts assessment in text-to-image generation,

    J. Wang, J. Hu, X. Ma, H. Ma, Y . Zeng, and X. Wei, “Magicmirror: A large-scale dataset and benchmark for fine-grained artifacts assessment in text-to-image generation,”arXiv preprint arXiv:2509.10260, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Where, what, why, and importance: Structured defect grounding for text-to-image feedback,

    H. Zhang, H. Yu, Y . Zhang, J. Wang, X. Chen, H. Cao, F. Lu, W. Zhang, C. Yu, and C. Yuan, “Where, what, why, and importance: Structured defect grounding for text-to-image feedback,”arXiv preprint arXiv:2606.06113, 2026

    Pith/arXiv arXiv 2026

  24. [24]

    A bias-free training paradigm for more general ai-generated image detection,

    F. Guillaro, G. Zingarini, B. Usman, A. Sud, D. Cozzolino, and L. Ver- doliva, “A bias-free training paradigm for more general ai-generated image detection,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 18 685–18 694

    2025

  25. [25]

    Qwen3-vl technical report,

    S. Bai, Y . Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Geet al., “Qwen3-vl technical report,”arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Deep residual injection for full-spectrum forensic signal perception in multimodal large language models,

    K. Lin, Z. Yan, R. Chen, K.-Y . Zhang, Y . Zhou, C. Piao, B. Li, T. Yao, B. Wang, Y . Xiaoet al., “Deep residual injection for full-spectrum forensic signal perception in multimodal large language models,”arXiv preprint arXiv:2606.15880, 2026

    arXiv 2026

  27. [27]

    Aegis: A holistic benchmark for evaluating forensic analysis of ai-generated academic images,

    B. Zhang, T.-Y . Ma, Z. Tang, J. Ding, Z. Wang, Y . Zhao, P. Gao, Z. Xi, Z. Ding, H. Sunet al., “Aegis: A holistic benchmark for evaluating forensic analysis of ai-generated academic images,”arXiv preprint arXiv:2604.28177, 2026

    Pith/arXiv arXiv 2026

  28. [28]

    Textshield-r1: Reinforced reasoning for tampered text detection,

    C. Qu, Y . Zhong, J. Liu, X. Zhu, B. Yu, and L. Jin, “Textshield-r1: Reinforced reasoning for tampered text detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 10, 2026, pp. 8621–8629

    2026

  29. [29]

    Fraudbench: A multimodal benchmark for detecting ai-generated fraudulent refund evidence,

    X. Yan, B. Chen, J. Zhang, T. Wu, H. X. Tae, Y . He, T. Wang, Y . Mi, Y . Hao, Y . Zhaoet al., “Fraudbench: A multimodal benchmark for detecting ai-generated fraudulent refund evidence,”arXiv preprint arXiv:2605.08820, 2026

    Pith/arXiv arXiv 2026

  30. [30]

    Beyond artifacts: Real-centric envelope modeling for reliable ai-generated image detection,

    R. Liu, Y . Han, Z. Zhang, L. Yao, Z. Yan, J. Shen, Z. Chen, B. Sun, L. Weng, J. Donget al., “Beyond artifacts: Real-centric envelope modeling for reliable ai-generated image detection,”arXiv preprint arXiv:2512.20937, 2025

    arXiv 2025

  31. [31]

    Bridging the gap between ideal and real-world evaluation: Benchmarking ai-generated image detection in challenging scenarios,

    C. Li, X. Wang, M. Li, B. Miao, P. Sun, Y . Zhang, X. Ji, and Y . Zhu, “Bridging the gap between ideal and real-world evaluation: Benchmarking ai-generated image detection in challenging scenarios,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 20 379–20 389

    2025

  32. [32]

    Genimage: A million-scale benchmark for detecting ai- generated image,

    M. Zhu, H. Chen, Q. Yan, X. Huang, G. Lin, W. Li, Z. Tu, H. Hu, J. Hu, and Y . Wang, “Genimage: A million-scale benchmark for detecting ai- generated image,”Advances in neural information processing systems, vol. 36, pp. 77 771–77 782, 2023

    2023

  33. [33]

    Wildfake: A large-scale and hierarchical dataset for ai-generated im- ages detection,

    Y . Hong, J. Feng, H. Chen, J. Lan, H. Zhu, W. Wang, and J. Zhang, “Wildfake: A large-scale and hierarchical dataset for ai-generated im- ages detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 4, 2025, pp. 3500–3508

    2025

  34. [34]

    Df40: Toward next-generation deepfake detec- tion,

    Z. Yan, T. Yao, S. Chen, Y . Zhao, X. Fu, J. Zhu, D. Luo, C. Wang, S. Ding, Y . Wuet al., “Df40: Toward next-generation deepfake detec- tion,”Advances in Neural Information Processing Systems, vol. 37, pp. 29 387–29 434, 2024

    2024

  35. [35]

    Community forensics: Using thousands of generators to train fake image detectors,

    J. Park and A. Owens, “Community forensics: Using thousands of generators to train fake image detectors,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 8245–8257

    2025

  36. [36]

    Forensichub: A unified benchmark & codebase for all-domain fake image detection and localization,

    B. Du, X. Zhu, X. Ma, C. Qu, K. Feng, Z. Yang, C.-M. Pun, J.-Z. Zhou et al., “Forensichub: A unified benchmark & codebase for all-domain fake image detection and localization,”Advances in Neural Information Processing Systems, vol. 38, 2026

    2026

  37. [37]

    A sanity check for ai-generated image detection,

    S. Yan, O. Li, J. Cai, Y . Hao, X. Jiang, Y . Hu, and W. Xie, “A sanity check for ai-generated image detection,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 70 702–70 720

    2025

  38. [38]

    So-fake: Benchmarking and explaining social media image forgery detection,

    Z. Huang, T. Li, X. Li, H. Wen, Y . He, J. Zhang, H. Fei, X. Yang, X. Huang, B. Penget al., “So-fake: Benchmarking and explaining social media image forgery detection,”arXiv preprint arXiv:2505.18660, 2025

    arXiv 2025

  39. [39]

    Is artificial intelligence generated image detection a solved problem?

    Z. Li, J. Yan, Z. He, K. Zeng, W. Jiang, L. Xiong, and Z. Fu, “Is artificial intelligence generated image detection a solved problem?”Advances in Neural Information Processing Systems, vol. 38, 2026. 14

    2026

  40. [40]

    Busterx: Mllm-powered ai-generated video forgery detection and explanation,

    H. Wen, Y . He, Z. Huang, T. Li, Z. Yu, X. Huang, L. Qi, B. Wu, X. Li, and G. Cheng, “Busterx: Mllm-powered ai-generated video forgery detection and explanation,”arXiv preprint arXiv:2505.12620, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    D3qe: Learning discrete distribution discrepancy-aware quantization error for autoregressive-generated image detection,

    Y . Zhang, B. Yu, Y . Zheng, W. Zheng, Y . Duan, L. Chen, J. Zhou, and J. Lu, “D3qe: Learning discrete distribution discrepancy-aware quantization error for autoregressive-generated image detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 16 292–16 301

    2025

  42. [42]

    No pixel left behind: A detail-preserving archi- tecture for robust high-resolution ai-generated image detection,

    L. Mu, Z. Xingze, J. Bai, J. Hu, W. Zheng, J. Ye, J. Zhuang, M. Ali, J. Wang, and H. Hu, “No pixel left behind: A detail-preserving archi- tecture for robust high-resolution ai-generated image detection,”arXiv preprint arXiv:2508.17346, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Next-imdl: Build benchmark for next-generation image manipulation detection & localization,

    Y . Li, H. He, Y . Zheng, B. Yu, W. Zheng, L. Chen, J. Zhou, and J. Lu, “Next-imdl: Build benchmark for next-generation image manipulation detection & localization,”arXiv preprint arXiv:2512.23374, 2025

    arXiv 2025

  44. [44]

    Gpt-image-2 in the wild: A twitter dataset of self-reported ai-generated images from the first week of deployment,

    K. Zewde, S. Ren, X. Shen, J. Wu, Y . Zhou, T. Duong, Z. Zhang, E. Traister, and K. Xie, “Gpt-image-2 in the wild: A twitter dataset of self-reported ai-generated images from the first week of deployment,” arXiv preprint arXiv:2604.25370, 2026

    Pith/arXiv arXiv 2026

  45. [45]

    Dˆ 3: scaling up deepfake detection by learning from discrepancy,

    Y . Yang, Z. Qian, Y . Zhu, O. Russakovsky, and Y . Wu, “Dˆ 3: scaling up deepfake detection by learning from discrepancy,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 23 850–23 859

    2025

  46. [46]

    Leveraging frequency analysis for deep fake image recogni- tion,

    J. Frank, T. Eisenhofer, L. Sch ¨onherr, A. Fischer, D. Kolossa, and T. Holz, “Leveraging frequency analysis for deep fake image recogni- tion,” inInternational conference on machine learning. PMLR, 2020, pp. 3247–3258

    2020

  47. [47]

    Thinking in frequency: Face forgery detection by mining frequency-aware clues,

    Y . Qian, G. Yin, L. Sheng, Z. Chen, and J. Shao, “Thinking in frequency: Face forgery detection by mining frequency-aware clues,” inEuropean conference on computer vision. Springer, 2020, pp. 86–103

    2020

  48. [48]

    Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning,

    C. Tan, Y . Zhao, S. Wei, G. Gu, P. Liu, and Y . Wei, “Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 5, 2024, pp. 5052–5060

    2024

  49. [49]

    Patchcraft: Exploring texture patch for efficient ai-generated image detection,

    N. Zhong, Y . Xu, S. Li, Z. Qian, and X. Zhang, “Patchcraft: Exploring texture patch for efficient ai-generated image detection,”arXiv preprint arXiv:2311.12397, 2023

    Pith/arXiv arXiv 2023

  50. [50]

    All patches matter, more patches better: Enhance ai-generated image detection via panoptic patch learning,

    Z. Yang, R. Chen, Z. Yan, K.-Y . Zhang, X. Fu, S. Wu, X. Shu, T. Yao, S. Ding, Z. Qinet al., “All patches matter, more patches better: Enhance ai-generated image detection via panoptic patch learning,”arXiv preprint arXiv:2504.01396, 2025

    arXiv 2025

  51. [51]

    Aligned datasets im- prove detection of latent diffusion-generated images,

    A. S. Rajan, U. Ojha, J. Schloesser, and Y . J. Lee, “Aligned datasets im- prove detection of latent diffusion-generated images,” inThe Thirteenth International Conference on Learning Representations

  52. [52]

    Reduce the artifacts bias for more generalizable ai-generated image detection,

    Y . Li, Y . Yang, Z. Tan, G. Li, Z. Lei, and W. Wang, “Reduce the artifacts bias for more generalizable ai-generated image detection,”arXiv preprint arXiv:2605.14486, 2026

    Pith/arXiv arXiv 2026

  53. [53]

    Co-spy: Combining semantic and pixel features to detect synthetic images by ai,

    S. Cheng, L. Lyu, Z. Wang, X. Zhang, and V . Sehwag, “Co-spy: Combining semantic and pixel features to detect synthetic images by ai,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 13 455–13 465

    2025

  54. [54]

    Omniaid: Decoupling semantic and artifacts for universal ai-generated image detection in the wild,

    Y . Guo, J. Ye, C. Zhang, H. Kang, H. Fu, C. He, and W. Li, “Omniaid: Decoupling semantic and artifacts for universal ai-generated image detection in the wild,”arXiv preprint arXiv:2511.08423, 2025

    Pith/arXiv arXiv 2025

  55. [55]

    Hy- draprompt: An adaptive and asymmetric framework of vision-language models for synthetic image detection,

    S. Shi, H. Tan, Z. Tan, S. Feng, A. Liu, S. Escalera, and J. Wan, “Hy- draprompt: An adaptive and asymmetric framework of vision-language models for synthetic image detection,”arXiv preprint arXiv:2605.26421, 2026

    Pith/arXiv arXiv 2026

  56. [56]

    Simplicity prevails: The emergence of generalizable aigi detection in visual foundation models,

    Y . Zhou, X. He, K. Lin, B. Fan, F. Ding, and B. Li, “Simplicity prevails: The emergence of generalizable aigi detection in visual foundation models,”arXiv preprint arXiv:2602.01738, 2026

    Pith/arXiv arXiv 2026

  57. [57]

    Mirror: Manifold ideal reference reconstructor for general- izable ai-generated image detection,

    R. Liu, M. Cui, Z. Qin, Z. Yan, R. Chen, Y . Han, Z. Li, J. Chen, Z. Chen, K. Linet al., “Mirror: Manifold ideal reference reconstructor for general- izable ai-generated image detection,”arXiv preprint arXiv:2602.02222, 2026

    arXiv 2026

  58. [58]

    Saido: Generalizable detection of ai-generated images via scene-aware and importance-guided dynamic optimization in continual learning,

    Y . Hu, Y . Cheng, Y . Zhang, Y . Xie, and Z. Yin, “Saido: Generalizable detection of ai-generated images via scene-aware and importance-guided dynamic optimization in continual learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 3876–3886

    2026

  59. [59]

    Generalizable and adaptive continual learning framework for ai- generated image detection,

    H. Wang, J. Lan, Y . Kang, H. Zhu, W. Wang, Z. Zhang, and S. Wang, “Generalizable and adaptive continual learning framework for ai- generated image detection,”IEEE Transactions on Multimedia, 2026

    2026

  60. [60]

    Are vlms seeing or just saying? uncovering the illusion of visual re- examination,

    C. Shi, C. Yang, Y . Wu, L. Jin, B. Shui, T. Berg-Kirkpatrick, and X. Ma, “Are vlms seeing or just saying? uncovering the illusion of visual re- examination,” inInt. Conf. Mach. Learn., 2026

    2026

  61. [61]

    Mirage: The illusion of visual understanding,

    M. Asadi, J. W. O’Sullivan, F. Cao, T. Nedaee, K. Rajabalifardi, F.-F. Li, E. Adeli, and E. Ashley, “Mirage: The illusion of visual understanding,” arXiv preprint arXiv:2603.21687, 2026

    arXiv 2026

  62. [62]

    Pixel reasoner: Incentivizing pixel space reasoning via curiosity-driven reinforcement learning,

    A. Su, H. Wang, W. Ren, F. Lin, and W. Chen, “Pixel reasoner: Incentivizing pixel space reasoning via curiosity-driven reinforcement learning,”Advances in Neural Information Processing Systems, vol. 38, pp. 8222–8251, 2026

    2026

  63. [63]

    Deepeyes: Incentivizing

    Z. Zheng, M. Yang, J. Hong, C. Zhao, G. Xu, L. Yang, C. Shen, and XingYu, “Deepeyes: Incentivizing ”thinking with images” via reinforce- ment learning,” inThe Fourteenth International Conference on Learning Representations, 2026

    2026

  64. [64]

    Mini-o3: Scaling up reasoning patterns and interaction turns for visual search,

    X. Lai, J. Li, W. Li, T. Liu, T. Li, and H. Zhao, “Mini-o3: Scaling up reasoning patterns and interaction turns for visual search,” inThe Fourteenth International Conference on Learning Representations, 2026

    2026

  65. [65]

    Thyme: Think beyond images,

    Y . Zhang, X. Lu, S. Yin, C. Fu, W. Chen, X. Hu, B. Wen, K. Jiang, C. Liu, T. Zhang, H. fan, K. Chen, J. Chen, H. Ding, K. Tang, Z. Zhang, L. Wang, F. Yang, T. Gao, and G. Zhou, “Thyme: Think beyond images,” inThe Fourteenth International Conference on Learning Representa- tions, 2026

    2026

  66. [66]

    Spotlight on token perception for multimodal reinforcement learning,

    S. Huang, X. Qu, Y . Li, Y . Luo, Z. He, D. Liu, and Y . Cheng, “Spotlight on token perception for multimodal reinforcement learning,” inThe Fourteenth International Conference on Learning Representations, 2026

    2026

  67. [67]

    Perception-aware policy optimization for multimodal reasoning,

    Z. Wang, X. Guo, S. Stoica, H. Xu, H. W ANG, H. Ha, X. Chen, Y . Chen, M. Yan, F. Huang, and H. Ji, “Perception-aware policy optimization for multimodal reasoning,” inThe Fourteenth International Conference on Learning Representations, 2026

    2026

  68. [68]

    Zooming without zooming: Region-to-image distillation for fine-grained multimodal perception,

    L. Wei, L. He, J. Lan, L. Dong, Y . Cai, S. Li, H. Zhu, W. Wang, L. Kong, Y . Wanget al., “Zooming without zooming: Region-to-image distillation for fine-grained multimodal perception,” inInternational Conference on Machine Learning, 2026

    2026

  69. [69]

    Agentic jigsaw interaction learning for enhancing visual perception and reasoning in vision-language models,

    Y . Zeng, W. Huang, S. Huang, X. Bao, Y . Qi, Y . Zhao, Q. Wang, L. Chen, Z. Chen, H. Chen, W. Ouyang, and F. Zhao, “Agentic jigsaw interaction learning for enhancing visual perception and reasoning in vision-language models,” inThe Fourteenth International Conference on Learning Representations, 2026

    2026

  70. [70]

    MiniLLM: Knowledge distillation of large language models,

    Y . Gu, L. Dong, F. Wei, and M. Huang, “MiniLLM: Knowledge distillation of large language models,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  71. [71]

    On-policy distillation of language models: Learning from self-generated mistakes,

    R. Agarwal, N. Vieillard, Y . Zhou, P. Stanczyk, S. R. Garea, M. Geist, and O. Bachem, “On-policy distillation of language models: Learning from self-generated mistakes,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  72. [72]

    Self-distillation enables continual learning,

    I. Shenfeld, M. Damani, J. H ¨ubotter, and P. Agrawal, “Self-distillation enables continual learning,”arXiv preprint arXiv:2601.19897, 2026

    Pith/arXiv arXiv 2026

  73. [73]

    Self-distilled reasoner: On-policy self-distillation for large language models,

    S. Zhao, Z. Xie, M. Liu, J. Huang, G. Pang, F. Chen, and A. Grover, “Self-distilled reasoner: On-policy self-distillation for large language models,”arXiv preprint arXiv:2601.18734, 2026

    Pith/arXiv arXiv 2026

  74. [74]

    Vision- opd: Learning to see fine details for multimodal llms via on-policy self- distillation,

    Q. Yuan, J. Lou, X. Yu, H. Lin, L. Sun, X. Han, and Y . Lu, “Vision- opd: Learning to see fine details for multimodal llms via on-policy self- distillation,”arXiv preprint arXiv:2605.18740, 2026

    Pith/arXiv arXiv 2026

  75. [75]

    Self-distilled rlvr,

    C. Yang, C. Qin, Q. Si, M. Chen, N. Gu, D. Yao, Z. Lin, W. Wang, J. Wang, and N. Duan, “Self-distilled rlvr,”arXiv preprint arXiv:2604.03128, 2026

    Pith/arXiv arXiv 2026

  76. [76]

    Asymmetric on-policy distillation: Bridging exploitation and imitation at the token level,

    N. Jia, H. Yang, X. Ma, J. Lian, S. Zhang, W. Zhang, K. Zeng, X. Cai, and Z. Sun, “Asymmetric on-policy distillation: Bridging exploitation and imitation at the token level,”arXiv preprint arXiv:2605.06387, 2026

    Pith/arXiv arXiv 2026

  77. [77]

    Entropy-aware on-policy distillation of language models,

    W. Jin, T. Min, Y . Yang, S. R. Kadhe, Y . Zhou, D. Wei, N. Baracaldo, and K. Lee, “Entropy-aware on-policy distillation of language models,” arXiv preprint arXiv:2603.07079, 2026

    Pith/arXiv arXiv 2026

  78. [78]

    Self-distilled agentic reinforcement learning,

    Z. Lu, Z. Yao, Z. Han, Z.-H. Wang, J. Wu, Q. Gu, X. Cai, W. Lu, J. Xiao, Y . Zhuanget al., “Self-distilled agentic reinforcement learning,”arXiv preprint arXiv:2605.15155, 2026

    Pith/arXiv arXiv 2026

  79. [79]

    Synartifact: Classifying and alleviating artifacts in synthetic images via vision- language model,

    B. Cao, J. Yuan, Y . Liu, J. Li, S. Sun, J. Liu, and B. Zhao, “Synartifact: Classifying and alleviating artifacts in synthetic images via vision- language model,”arXiv preprint arXiv:2402.18068, 2024

    Pith/arXiv arXiv 2024

  80. [80]

    Unveiling perceptual artifacts: A fine- grained benchmark for interpretable AI-generated image detection,

    Y . Xiao, W. Chen, J. Chen, Z. Cao, W. Deng, B. Yang, Z. Dong, X. Ji, W. Ke, P. Wei, and L. Lin, “Unveiling perceptual artifacts: A fine- grained benchmark for interpretable AI-generated image detection,” in The Fourteenth International Conference on Learning Representations, 2026

    2026

Showing first 80 references.