Pith. sign in

REVIEW 4 major objections 5 minor 21 references

CMA quantifies, per prediction, how much the image, the text, and their joint interaction contribute to a multimodal LLM's decision; on controlled tests it names the decision-driving modality in 98% of cases.

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 · deepseek-v4-flash

2026-08-05 04:20 UTC pith:DYBBKGEW

load-bearing objection Promising framework for modality-level attribution, but the headline accuracy rests on easy biased classifiers and the method has two under-specified steps; worth reviewing, needs major revision. the 4 major comments →

arxiv 2608.00076 v2 pith:DYBBKGEW submitted 2026-07-29 cs.CV cs.AI

Which Modality Decides? Counterfactual Modality Attribution for Multimodal LLMs

Vahidin Hasic , Chao Wang , Luis C. Garcia-Peraza-Herrera , David Watson , Senka Krivic This is my paper
classification cs.CV cs.AI
keywords modality attributioncounterfactual explanationsmultimodal large language modelsShapley valuesdiffusion modelsexplainabilitycross-modal reasoningmedical imaging
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.

The paper argues that existing explainability tools answer where a model looks—which pixels or tokens matter—but not which whole channel, image or text, actually decides a prediction. It introduces CMA, a counterfactual modality-attribution framework: the method generates plausible edited versions of the image alone, the text alone, and both together, then evaluates the classifier on the four resulting input coalitions and converts the four logit values into Shapley-based Image% and Text% scores. On synthetic benchmarks with known ground truth, CMA attributes the correct modality 98% of the time and beats feature-, perturbation-, and surrogate-based baselines. On a real chest X-ray benchmark, it exposes a model that diagnoses a nodule from ECG leads in the image rather than from radiographic evidence—an unsafe reliance that prediction accuracy alone cannot reveal. The paper's broader claim is that modality attribution should be treated as a separate, complementary explainability objective for multimodal foundation models.

Core claim

The central claim is that a multimodal classifier's prediction can be decomposed, at the level of whole input channels, into an image contribution, a text contribution, and their interaction, by measuring how the logit of the predicted class changes under realistic counterfactual edits of each modality. CMA computes this decomposition by generating three counterfactual inputs—image-only, text-only, and joint—through synchronized reverse diffusion guided by classifier gradients, then evaluates the classifier at the four combinations of original and edited modalities. The closed-form two-player Shapley values of this game give the modality attribution scores, and the normalized absolute scores

What carries the argument

The central object is a two-player counterfactual coalition grid: for an input pair (image, text), CMA creates an image-only counterfactual, a text-only counterfactual, and a joint counterfactual, so that the classifier logit for the original predicted class can be evaluated on all four coalitions {both original, image edited, text edited, both edited}. These values feed the closed-form two-player Shapley formula, which averages marginal contributions and isolates cross-modal interaction. The generating machinery is synchronized reverse diffusion: a Denoising Diffusion Probabilistic Model (DDPM) for images and a Riemannian Diffusion Language Model (RDLM) for text move through reverse time at

Load-bearing premise

The accuracy claim rests on the untested premise that the synchronized image and text diffusion processes produce realistic, comparable counterfactuals; if the guiding prior drifts off the data manifold, the classifier logits used in the Shapley game are out-of-distribution and the attribution scores become unreliable.

What would settle it

Take the chest X-ray case where CMA attributes the prediction to an image shortcut: manually remove the ECG leads from the image while leaving the radiology report identical, and run the classifier. If the model still predicts the nodule class, CMA's attribution to the image shortcut is wrong.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Audits of high-stakes multimodal systems can report a modality split (Image% vs Text%) alongside accuracy, so a correct-but-shortcutting model becomes visible before deployment.
  • Modality attribution accuracy could become a benchmark metric: two models with equal task accuracy can have opposite modality reliance, so accuracy alone is not a sufficient evaluation target.
  • Joint counterfactuals provide a measurable test of genuine multimodal fusion, since their effect on the prediction is not the sum of the two unimodal edits.
  • In clinical settings, CMA-style attribution can flag cases where a diagnosis is supported by spurious image cues, prompting re-labeling or re-training on the real pathological evidence.
  • Because CMA is model-agnostic and only requires differentiability, the same four-coalition Shapley grid applies to any multimodal architecture, from document understanding to robotics.

Where Pith is reading between the lines

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

  • Going beyond the paper, modality attribution could be used as a training signal: downweighting samples where the attributed dominant modality contradicts the task's intended evidence might reduce shortcut reliance without extra labels.
  • The two-player game naturally extends to three or more input channels (image, structured records, free text), although that would require synchronized counterfactual priors for every new channel—an untested generalization.
  • Because CMA is reference-dependent, a fair model-to-model comparison would need to fix the counterfactual reference distribution; otherwise attribution gaps may reflect generator differences rather than behavioral differences.
  • A hierarchical extension is plausible: once the dominant modality is identified, running the same game inside that modality could locate the specific image regions or text spans that drive the effect, unifying modality attribution with feature attribution.

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

4 major / 5 minor

Summary. The paper introduces Counterfactual Modality Attribution (CMA), a post-hoc explainability method for multimodal LLMs that quantifies the contribution of image and text modalities to a classifier's prediction. CMA generates image-only, text-only, and joint counterfactuals by running synchronized DDPM and RDLM diffusion priors under classifier guidance, then computes two-player Shapley values using the logit of the original predicted class on the four coalitions. The method is evaluated on controlled synthetic MNIST variants with known modality biases and on the OpenI chest X-ray dataset. The authors report 98% attribution accuracy on the biased classifiers and demonstrate a qualitative clinical shortcut (ECG leads) on OpenI.

Significance. CMA addresses a real gap: existing feature-attribution methods do not answer which modality drives an MLLM prediction. The formulation is principled, and the use of counterfactual interventions rather than gradients or static perturbations is well motivated. The paper's strengths include the controlled synthetic benchmarks with known ground truth, the inclusion of source code in the supplement, the closed-form two-player Shapley solution, and a corruption-based sanity check. However, as presented, the central 98% accuracy claim rests on two biased synthetic classifiers and on a counterfactual generation pipeline whose synchronization and text update rule are under-specified. If these issues are resolved, CMA would be a valuable contribution to multimodal explainability.

major comments (4)
  1. [Methodology: Coupled Diffusion] The claimed synchronization between DDPM (discrete step s) and RDLM (continuous time t) is not defined. The text says the reverse processes are 'synchronized at the same diffusion timestep' and Fig. 2 says 'both arms same step t', but no mapping is given between s and t, nor is the number of reverse steps or the alignment schedule specified. Because the joint counterfactual is generated by conditioning each modality on the other's evolving state, different alignments change the joint trajectory and therefore v(∅) and the resulting Shapley values. The 98% accuracy claim depends on this unspecified choice. Please provide the exact coupling mechanism (e.g., time reparameterization, step-count matching) and a sensitivity analysis over alignment choices.
  2. [Methodology: Text-only counterfactuals] The update pθ ← softmax(log pθ − (λ_c^T ĝ^T + λ_1^T (pθ − 1_{d0}))) contains the undefined symbol `1 d0`. If it is not the exact original token representation (one-hot, embedding, or predictive distribution), the proximal penalty is not a well-defined function and the text counterfactual may be neither minimal nor on-manifold. Also, ĝ^T = ∇_{ℓθ} L_CW is computed w.r.t. per-digit logits while the update acts on probabilities; the chain rule is not shown. Please define all symbols and provide the exact update used in the experiments.
  3. [Results: Synthetic Multimodal MNIST] The headline 98% accuracy is reported only for the Image-Biased and Text-Biased classifiers, where the ground-truth decision rule is degenerate. In these settings any method that directly tests whether replacing one modality changes the prediction is predisposed to succeed; the result does not establish that CMA resolves the Balanced variant, which requires both modalities and is the case where modality attribution is nontrivial. The Balanced variant appears only in the corruption ablation (Fig. 5b), not in the attribution-accuracy comparison (Fig. 4). Please report attribution accuracy and interaction measures on the Balanced variant, or justify why biased classifiers suffice.
  4. [Limitations] The paper concedes that 'imperfect counterfactual generation may affect attribution quality', but provides no experimental support for the adequacy of the generated counterfactuals. There is no report of flip success rates per generation mode, no distance-to-original statistics, no human or automated realism evaluation, and no sensitivity analysis over the guidance weights λ_c, λ_1, the inpainting threshold τ_img, or the token-change threshold τ_txt. Since every attribution score in the paper is computed from these counterfactuals, the absence of such validation is load-bearing for the central claim.
minor comments (5)
  1. [Results: Multiclass OpenI] The OpenI evaluation is qualitative only. The paper reports one illustrative counterfactual (ECG leads removed) but no quantitative metrics such as flip rates, attribution distributions, or agreement with clinical labels. Please consider adding such numbers to support the claim that CMA reveals clinically relevant shortcuts.
  2. [Methodology: Image-only counterfactuals] The notation `sign(·)` and the inpainting threshold τ_img are not numerically specified. Please provide values or point to the supplementary material for all hyperparameters used in the experiments.
  3. [Background: Text Diffusion] The RDLM notation uses X_t^T for both token representation and time, and the drift expression is dense. Please clarify the variable roles and define all symbols (e.g., e_k, c(t), β(t)) in one place.
  4. [Results: Figure 4] The text reports exact hit rates only for MM-SHAP and MultiViz. Please report numerical values with confidence intervals for all baselines in the figure or caption, so the reader can verify the 'consistently outperforms' claim.
  5. [Introduction] The phrase 'first framework' is used in the abstract and contributions. Given the rapid growth of multimodal explainability work, please temper this claim or add a more thorough comparison to recent modality-attribution methods, especially those using causal or counterfactual interventions.

Circularity Check

0 steps flagged

No significant circularity: CMA's attribution is an intervention-based Shapley measure validated on externally defined ground-truth benchmarks; the cited self-reference is not load-bearing.

full rationale

No circular steps found. CMA computes modality attribution from classifier logits on diffusion-generated counterfactuals, and its synthetic benchmarks define ground-truth modality reliance by construction of the data-generation process (Image-Biased and Text-Biased variants), independent of CMA's parameters or outputs. The 98% hit rate is therefore an empirical validation of the intervention logic, not a fitted quantity renamed as a prediction. The only self-citation (Wang et al. 2026) appears in the related-work discussion of video counterfactuals and is not load-bearing for any central claim. The paper's own Limitations section explicitly acknowledges that imperfect counterfactual generation may affect attribution quality, which is a correctness risk rather than a circularity. The under-specified DDPM/RDLM time synchronization and the ambiguous `1_d0` token-representation term are implementation-level concerns that could affect the validity of the counterfactuals, but they do not make the attribution equal to its input by construction. The Shapley formulation is standard and the coalition values are computed directly from the pretrained classifier's logits on the original and counterfactual inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

CMA defines a new attribution score based on counterfactual interventions and Shapley values. The main free parameters are the counterfactual generation hyperparameters (guidance and proximity weights, change thresholds), which are not specified in the main text. The axioms are the modeling assumptions that make the counterfactuals and the Shapley game a valid measure of modality reliance.

free parameters (6)
  • Image guidance weight lambda_c^I = not stated (supplementary)
    Scales the classifier gradient in the image reverse diffusion update; affects how aggressively the image is changed to flip the prediction.
  • Image proximity weight lambda_1^I = not stated (supplementary)
    Penalizes deviation from the original image; controls minimality of the image counterfactual.
  • Text guidance weight lambda_c^T = not stated (supplementary)
    Scales the classifier gradient in the text diffusion update.
  • Text proximity weight lambda_1^T = not stated (supplementary)
    Penalizes deviation from the original token distribution.
  • Inpainting change threshold tau_img = not stated
    Restores pixels whose predicted changes are below threshold to preserve unchanged regions.
  • Token change threshold tau_txt = not stated
    Keeps tokens whose predicted change is below threshold unchanged.
axioms (4)
  • domain assumption The two diffusion priors (DDPM for images, RDLM for text) generate realistic counterfactual samples that lie on the data manifold.
    Invoked in Methodology: Counterfactual Generation; if counterfactuals are off-manifold, the classifier logits are out-of-distribution and Shapley values become unreliable.
  • domain assumption The Shapley value computed on logit differences over these four coalitions captures the causal modality contribution to the prediction.
    The paper equates modality attribution with this cooperative game (Methodology: Modality Attribution); this is a modeling choice, not a proven theorem.
  • domain assumption For the synthetic benchmarks, the ground-truth modality reliance is defined by the training setup, e.g., a model trained to predict only the digit is assumed to rely only on the image.
    Used to define hit rate; this is a reasonable experimental design assumption.
  • domain assumption Gradients from the pretrained MLLM classifier can be back-propagated through the soft text embeddings and the image input to guide both diffusion processes coherently.
    Required during counterfactual generation; the paper does not demonstrate that the combined gradient path is stable.

pith-pipeline@v1.3.0-alltime-deepseek · 12276 in / 13436 out tokens · 135553 ms · 2026-08-05T04:20:55.896032+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Which Modality Decides? Counterfactual Modality Attribution for Multimodal LLMs." pith.science (2026). https://pith.science/paper/DYBBKGEW

@misc{pith2026260800076,
  author       = {Pith},
  title        = {Pith review of: Which Modality Decides? Counterfactual Modality Attribution for Multimodal LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DYBBKGEW}},
  note         = {Machine review of arXiv:2608.00076}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multimodal large language models (MLLMs) increasingly support high-stakes decision making by combining complementary information from images and text. While existing explainability methods identify influential image regions or text tokens, they cannot answer a fundamental question: which modality drives a prediction? Consequently, a model may produce the correct output while relying on the wrong source of evidence, masking shortcut learning and unsafe reasoning. We formulate modality attribution as a complementary explainability objective for multimodal foundation models and propose Counterfactual Modality Attribution (CMA), the first framework for quantifying modality-level contributions in MLLMs. CMA generates image-only, text-only, and joint multimodal counterfactuals using coupled diffusion priors and converts them into principled modality attribution scores through a cooperative game-theoretic formulation based on Shapley values. We evaluate CMA on controlled synthetic benchmarks with known ground-truth modality reliance and on a real-world multimodal clinical dataset. CMA correctly identifies the decision-driving modality in 98% of controlled cases and consistently outperforms baselines, revealing failures of cross-modal reasoning that remain invisible to predictive accuracy alone. Our results establish modality attribution as a complementary dimension of explainability beyond feature attribution, providing a principled framework for auditing multimodal foundation models in safety-critical applications.

Figures

Figures reproduced from arXiv: 2608.00076 by Chao Wang, David Watson, Luis C. Garcia-Peraza-Herrera, Senka Krivic, Vahidin Hasic.

Figure 1
Figure 1. Figure 1: Illustration of modality attribution failure in a medi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed Counterfactual Modality Attribution (CMA) framework. (1) Independently pretrained image [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of generated counterfactuals across different generation modes for Synthetic MNIST (top row, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Modality attribution accuracy on image-biased and [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Modality attribution share (%) under progressive corruption of image and text modalities for (a) OpenI medical dataset and (b) Synthetic MNIST dataset. Error bands indicate standard error across test samples. As corruption increases, attribution share decays to 0%, shifting entirely to the uncorrupted modality. demonstrates that CMA is input-faithful and adapts its at￾tribution to the evidence available fo… view at source ↗

Review history (2 revisions) →

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

21 extracted references · 9 canonical work pages

  1. [5]

    arXiv preprint arXiv:2412.02104

    Explainable and Interpretable Multi- modal Large Language Models: A Comprehensive Survey. arXiv preprint arXiv:2412.02104. Demner-Fushman, D.; Kohli, M. D.; Rosenman, M. B.; Shooshan,S.E.;Rodriguez,L.;Antani,S.;Thoma,G.R.;and McDonald, C. J

    Pith/arXiv arXiv

  2. [8]

    Gemma Team

    Scal- ing and Evaluating Sparse Autoencoders.arXiv preprint arXiv:2406.04093. Gemma Team

    Pith/arXiv arXiv

  3. [9]

    Guidotti, R

    Gemma 3 Technical Report.arXiv preprint arXiv:2503.19786. Guidotti, R

    Pith/arXiv arXiv

  4. [11]

    InAdvances in Neural Information Processing Systems

    Continuous Diffusion Model for Language Modeling. InAdvances in Neural Information Processing Systems. Joshi, S.; Yin, H.; Adiga, R.; Monti, R.; Carranza, A.; Fang, A.;Deng,A.;Abbas,A.;Larsen,B.;Blakeney,C.;etal.2026. DatBench:Discriminative,Faithful,andEfficientVLMEval- uations.arXiv preprint arXiv:2601.02316. LeCun, Y.; Bottou, L.; Bengio, Y.; Haffner, ...

    arXiv 2026

  5. [14]

    Lou, H.; Li, C.; Ji, J.; and Yang, Y

    A Survey on Mechanistic Interpretabil- ity for Multi-Modal Foundation Models.arXiv preprint arXiv:2502.17516. Lou, H.; Li, C.; Ji, J.; and Yang, Y

    Pith/arXiv arXiv

  6. [15]

    InInternationalConferenceonLearningRepresen- tations

    Sparse Feature Circuits: Discover- ing and Editing Interpretable Causal Graphs in Language Models. InInternationalConferenceonLearningRepresen- tations. Nguyen,V.B.;Seifert,C.;andSchlötterer,J.2025. Guiding LLMstoGenerateHigh-FidelityandHigh-QualityCounter- factual Explanations for Text Classification.arXiv preprint arXiv:2503.04463. Nie, S.; Zhu, F.; You...

    work page arXiv 2025

  7. [16]

    Olson, M

    Large Language Diffusion Models.arXiv preprint arXiv:2502.09992. Olson, M. L.; Hinck, M.; Ratzlaff, N.; Li, C.; Howard, P.; Lal,V.;andTseng,S.-Y.2025. ProbingtheRepresentational Power of Sparse Autoencoders in Vision Models.arXiv preprint arXiv:2508.11277. Parcalabescu, L.; and Frank, A

    Pith/arXiv arXiv 2025

  8. [17]

    In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 3094–3109

    MiCE: ExplainingNLPModelsviaMinimalContrastiveEditing. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 3094–3109. Selvaraju, R. R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; and Batra, D

    work page 2021

  9. [18]

    InProceedingsoftheIEEEinternationalconferenceoncom- puter vision, 618–626

    Grad-cam: Visual expla- nations from deep networks via gradient-based localization. InProceedingsoftheIEEEinternationalconferenceoncom- puter vision, 618–626. Shapley,L.S.1953. Avalueforn-persongames. InContri- butionstotheTheoryofGames,volume2,307–317.Prince- ton University Press. Simon, P.; et al

    work page 1953

  10. [19]

    Soenksen, L

    The future of multimodal AI in medicine: a pragmatic roadmap.arXiv preprint arXiv:2505.20316. Soenksen, L. R.; Ma, Y.; Zeng, C.; Boussiarès, L.; Ting, D. S. W.; Rajpurkar, P.; and Shah, N. H

    Pith/arXiv arXiv

  11. [20]

    InEuropean Conference on Computer Vision, 351–368

    Fast Diffusion-Based Counterfactuals for Short- cut Removal and Generation. InEuropean Conference on Computer Vision, 351–368. Springer. Wu,T.;Ribeiro,M.T.;Heer,J.;andWeld,D.2021.Polyjuice: Generating Counterfactuals for Explaining, Evaluating, and ImprovingModels. InProceedingsofthe59thAnnualMeet- ing of the Association for Computational Linguistics, 6707–

    work page 2021

  12. [565]

    Omni-Diffusion:UnifiedMulti- modal Understanding and Generation with Masked Discrete Diffusion

    Li, L.; Long, Z.; Shen, Y.; Gao, H.; Cao, H.; Sun, X.; Shan, C.;He,R.;andFu,C.2026. Omni-Diffusion:UnifiedMulti- modal Understanding and Generation with Masked Discrete Diffusion. InInternational Conference on Machine Learn- ing. Li,Y.;etal.2025. TokenActivationMaptoVisuallyExplain Multimodal LLMs. InIEEE/CVF International Conference on Computer Vision. L...

    work page 2026

  13. [2017]

    Asadi, M.; O’Sullivan, J

    Towards better understanding of gradient-based attribu- tion methods for deep neural networks.arXiv preprint arXiv:1711.06104. Asadi, M.; O’Sullivan, J. W.; Cao, F.; Nedaee, T.; Raja- balifardi, K.; Li, F.-F.; Adeli, E.; and Ashley, E

    Pith/arXiv arXiv

  14. [2020]

    InAdvances in Neural Information Processing Systems, volume 33, 6840–6851

    Denoising Diffusion Probabilistic Models. InAdvances in Neural Information Processing Systems, volume 33, 6840–6851. Jeanneret,G.;Simon,L.;andJurie,F.2022. Diffusionmod- els for counterfactual explanations. InProceedings of the Asian conference on computer vision, 858–876. Jeanneret, G.; Simon, L.; and Jurie, F

    work page 2022

  15. [2021]

    Dunefsky,J.;Chlenski,P.;andNanda,N.2024

    Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34: 8780–8794. Dunefsky,J.;Chlenski,P.;andNanda,N.2024. Transcoders Find Interpretable LLM Feature Circuits. InAdvances in Neural Information Processing Systems. Fu, Y.; Whalen, L.; Ye, Z.; Dong, X.; Diao, S.; Liu, J.; Wu, C.; Zhang, H.; Xie, E.; Han, S.; et al

    work page 2024

  16. [2022]

    arXiv preprint arXiv:2207.00056

    Multiviz: Towards visualizing and understanding multimodal models. arXiv preprint arXiv:2207.00056. Lin, Y.; et al

    Pith/arXiv arXiv

  17. [2023]

    Ben Melech Stan, G.; Aflalo, E.; Bhatt, G.; Chen, S.-Y

    DiG- IN: Diffusion Guidance for Investigating Networks– Uncovering Classifier Differences Neuron Visualisations and Visual Counterfactual Explanations.arXiv preprint arXiv:2311.17833. Ben Melech Stan, G.; Aflalo, E.; Bhatt, G.; Chen, S.-Y. T.; Tseng, S.-Y.; Sod-Moriah, G.; Tsiper, S.; Yaniv, Y.; and Lal, V

    Pith/arXiv arXiv

  18. [2024]

    InIEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 8182–8187

    LVLM-Interpret: An Interpretability Tool for Large Vision-Language Models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 8182–8187. Boreiko, V.; Augustin, M.; Croce, F.; Berens, P.; and Hein, M.2022. Sparsevisualcounterfactualexplanationsinimage space. InDAGM German Conference on Pattern Recogni- tion, 133–148. Springer. Carl...

    work page 2022

  19. [2025]

    Gao, L.; la Tour, T

    Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speed.arXiv preprint arXiv:2512.14067. Gao, L.; la Tour, T. D.; Tillman, H.; Goh, G.; Troll, R.; Radford, A.; Sutskever, I.; Leike, J.; and Wu, J

    Pith/arXiv arXiv

  20. [2026]

    Augustin, M.; Boreiko, V.; Croce, F.; and Hein, M

    MI- RAGE:TheIllusionofVisualUnderstanding.arXivpreprint arXiv:2603.21687. Augustin, M.; Boreiko, V.; Croce, F.; and Hein, M

    arXiv

  21. [6723]

    InProceedings of the 64th Annual Meeting of the Asso- ciation for Computational Linguistics

    Zhong, L.; Wu, L.; Fang, B.; Feng, T.; Jing, C.; Wang, W.; Zhang,J.;Chen,H.;andShen,C.2026.BeyondHardMasks: Progressive Token Evolution for Diffusion Language Mod- els. InProceedings of the 64th Annual Meeting of the Asso- ciation for Computational Linguistics

    work page 2026