Pith. sign in

REVIEW 4 major objections 5 minor 21 references

A training-free decoding policy that vetoes tokens with unstable confidence trajectories or weak visual grounding consistently improves reasoning accuracy in diffusion-based multimodal LLMs, by up to 9 percentage points, without extra train

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

2026-08-01 16:43 UTC pith:OBERZTN4

load-bearing objection Training-free veto-and-swap for diffusion MLLMs is a plausible new idea, but the main numbers need error bars, code, and a stronger control before I'd trust the mechanistic claim. the 4 major comments →

arxiv 2607.17884 v1 pith:OBERZTN4 submitted 2026-07-20 cs.AI

ST-Veto: Spatio-Temporal Token Veto for Diffusion MLLMs via Taylor Prediction and Visual Grounding

Keuntae Kim , Beomseok Lee , Hyunwoo Kim , Yong Suk Choi This is my paper
classification cs.AI
keywords diffusion multimodal LLMsdecoding policytoken vetoTaylor predictionvisual groundingimage attentiontraining-freechain-of-thought reasoning
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.

ST-Veto claims that the reasoning failures of diffusion-based multimodal large language models (dMLLMs) can be substantially reduced by a training-free decoding policy: before a token is committed as context for the next denoising step, check whether its confidence trajectory is smooth and whether it is visually grounded. The paper proposes two veto signals—a second-order Taylor residual that flags tokens whose confidence deviates from predicted dynamics, and an image-attention mass that flags weakly grounded tokens—and swaps vetoed tokens with safe near-boundary candidates. Across two dMLLMs and four benchmarks, the method improves accuracy by up to 9 percentage points at no extra generation cost. The result matters because it suggests that better reasoning in diffusion models can come from decoding-time control rather than retraining, and that autoregressive chain-of-thought methods do not transfer to the parallel denoising paradigm.

Core claim

ST-Veto establishes that in diffusion-based multimodal language models, the standard practice of unmasking the currently most confident tokens at each denoising step is suboptimal: tokens can be confidently wrong because they are locally obvious under a noisy context, yet unstable over time or poorly anchored to the image. The paper shows that a veto-and-swap policy, which rejects tokens with large negative Taylor residuals of confidence and tokens with low image-attention mass, and replaces them with safe candidates that are both temporally stable and visually grounded, consistently improves reasoning accuracy across models and benchmarks. The gains are asymmetric in paired analysis (many m

What carries the argument

The load-bearing object is the Veto-and-Swap policy. It defines a veto criterion from two signals: the Taylor residual, the deviation of a token's current confidence from a second-order Taylor prediction computed from a four-step history of that same masked position, and the image attention mass, the token's attention toward image tokens averaged over layers L/3, L/2, and 2L/3. A token is vetoed if either signal falls below median minus one MAD (median absolute deviation); a swap candidate is safe only if both signals exceed median plus one MAD and its confidence lies within one MAD of the top-k boundary. The policy then replaces vetoed top-k tokens with the highest-confidence safe candidate

Load-bearing premise

The whole veto signal rests on the premise that a token's model confidence and its attention to image tokens are trustworthy indicators of whether the token is correct and visually grounded; the paper itself concedes that both are imperfect proxies that can be misled by non-smooth confidence trajectories and diffuse or semantically misaligned image attention.

What would settle it

Run ST-Veto on a dMLLM where calibration is known to be poor—for example, one whose high-confidence tokens are no more likely correct than low-confidence ones—and compare accuracy against base decoding. If accuracy does not improve or degrades, the confidence-based veto signal is not doing the work. A more direct test: use a benchmark where the model is systematically overconfident about hallucinated objects and check whether the flip ratio becomes symmetric or negative.

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

If this is right

  • If correct, training-free decoding policies can improve dMLLM reasoning across benchmarks without retraining or extra generation steps.
  • Autoregressive chain-of-thought prompting methods are structurally mismatched with diffusion decoding and should not be expected to transfer; diffusion-native control is needed.
  • The veto signals track token-level competence: vetoed tokens are disproportionately function words, swapped tokens content words, supporting an early-anchor interpretation.
  • Gains persist at longer generation lengths and extend to additional benchmarks such as MME, POPE, and MMMU.
  • Temporal stability and visual grounding are complementary signals; combining them outperforms either alone in ablations.

Where Pith is reading between the lines

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

  • The same veto-and-swap principle could be tested on text-only discrete diffusion LLMs, where only the Taylor residual signal is available; a temporal-only veto might already improve coherence.
  • The second-order Taylor prediction can be viewed as a null model for smooth confidence evolution; its residual measures local volatility that could be linked to token-level uncertainty in future work.
  • ST-Veto's conservative design—swapping only when a safe near-boundary candidate exists—bounds its downside, making it a plausible default decoding strategy for production dMLLMs.
  • Combining ST-Veto with training-time objectives that explicitly smooth confidence trajectories or sharpen image grounding could yield further gains beyond decoding-time control.

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 proposes ST-Veto, a training-free decoding-time method for diffusion-based multimodal LLMs (dMLLMs). At each unmasking step, ST-Veto identifies top-k confident tokens and vetoes those whose confidence trajectory deviates from a second-order Taylor prediction (temporal instability) or whose image-attention mass falls below a MAD-based threshold (weak visual grounding). Vetoed tokens are replaced by safe near-boundary candidates that satisfy both temporal and grounding criteria. Experiments on M3CoT, ScienceQA, MMBench, and V*Bench with LaViDa and MMaDA claim consistent improvements over standard decoding baselines and adapted AR-style reasoning methods, with gains up to about 8.5 percentage points and negligible throughput overhead. Ablations, flip analysis, stagewise swap distributions, and POS analysis are presented to support the mechanism.

Significance. If the empirical claims hold, ST-Veto would be a valuable, practical contribution: a training-free, near-zero-cost decoding policy that improves reasoning in a class of models (dMLLMs) for which AR-style prompting methods do not transfer. The paper provides several internal analyses—ablation of each signal, stagewise swap distribution, instance-level flip asymmetry, POS distribution of vetoed/swapped tokens—that go beyond a single accuracy number. These analyses, together with the clear formulation of the veto-and-swap policy, make the mechanism plausible and the paper worth serious consideration. However, the central empirical claim currently rests on a few percentage-point gains reported without variance, and the mechanistic interpretation is not fully identified against a simpler 'swap-in-content-words' alternative.

major comments (4)
  1. [§5.1, Table 1] Table 1 reports averages over three seeds but no standard deviations, confidence intervals, or significance tests. Many improvements are small in absolute terms (e.g., MMaDA M3CoT 128/64: 26.32→27.52; MMaDA MMBench 256/128: 32.04→36.74 is larger but still needs variance). Without error bars, the reader cannot assess whether the single-digit gains are robust across seeds. Please report per-seed results or variance estimates, and ideally paired significance tests.
  2. [§5.3 (Table 3) vs Table 1] There is a load-bearing inconsistency: MMaDA MMBench 128/64 is 36.55 in Table 1 but 35.55 in Table 3, and 256/128 is 36.74 vs 35.74. Since Table 3 is the ablation of the main result, the numbers must agree. Clarify which table is correct and correct the typo; this discrepancy calls into question the reliability of the numerical reporting.
  3. [§3.3, §5.3, §6 (missing negative control)] The causal claim is that Taylor residuals and image-attention masses carry information about correctness and grounding, and that this information drives the gains. The paper ablates each signal off, but both ablations still use the veto-and-swap policy with some signal. Given that only 2–7% of top-k tokens are swapped (Table 7) and swapped tokens are predominantly content words (Table 13), the gains could in principle come almost entirely from the swap action itself—replacing function words with near-boundary content words—rather than from the specific information in e_i and m_i. A permutation control is needed: shuffle e_i and m_i across positions while preserving the veto rate, swap counts, and candidate-line thresholds. If permuted signals retain most of the gain, the mechanistic interpretation in Eqs. (5)–(14) is underdetermined.
  4. [§A.9, Eq. (11)–(12)] The method's premise is that internal confidence and attention are reliable proxies for token correctness and visual grounding. The paper acknowledges in A.9 that these are 'imperfect proxies,' but it never tests the premise directly. For instance, one could measure whether vetoed tokens are indeed less likely to appear in the final correct sequence, or whether high-confidence tokens correlate with answer correctness in these dMLLMs. Without such a check, the veto criteria (Eq. 11–12) are justified only by the end-task accuracy, which is circular if the proxy assumption itself is part of the claimed mechanism.
minor comments (5)
  1. [§3.3] MAD is used throughout (Eq. 11, 12, 14) but never defined. Define median absolute deviation and specify whether the scaling factor 1.4826 is used.
  2. [Figure 3] Panels (c) and (d) contain unreadable glyph sequences (e.g., '/uni00000018/...'), likely a font-encoding artifact. The figure should be regenerated with proper text rendering.
  3. [Table 14] The 'Attribute recognition' row has missing spacing ('26435.61') and should read '264 35.61'.
  4. [References] Sahoo et al. (2024a) and (2024b) are the same paper and should be merged or distinguished by venue details; currently listed twice with identical titles.
  5. [§6.2, Table 4] The text says H=4 'provides the best cost-performance trade-off,' but Table 4 shows the third-order variant (H=5) matching or exceeding H=4 in several cells (e.g., MMBench 128/64: 38.11 vs 38.05). The text acknowledges this, but the wording 'best' is slightly strong; consider 'best trade-off' or 'strongest overall.'

Circularity Check

1 steps flagged

Minor self-referential grounding analysis; central accuracy claims are independent.

specific steps
  1. self definitional [Section 6.3 / Figure 3; Eqs. (8)-(9) and (12)-(14)]
    "The results show that ST-Veto consistently achieves higher image attention ratios than the baseline across all steps. This can be interpreted as the effect of vetoing certain tokens that, despite having high confidence scores, exhibit weak visual grounding."

    The grounding signal m_i is defined in Eqs. (8)-(9) as image-attention mass, and C_safe in Eq. (12) admits swap candidates only if m_i >= median(m)+MAD(m). Thus every token admitted by the swap rule is, by construction, above-median in the same m_i that Figure 3 averages. Reporting higher average image attention under ST-Veto therefore restates the selection rule; it is not an independent confirmation that the chosen tokens are semantically better-grounded. This is a supporting analysis, not the source of the external benchmark gains in Table 1.

full rationale

The central claim of the paper—consistent accuracy improvements on M3CoT, ScienceQA, MMBench, and V*Bench—is validated against external benchmark answers, not against ST-Veto's own confidence or attention signals. No fitted constant or internal statistic is reused as the benchmark target: the Taylor residual (Eqs. 5-6) is a finite-difference extrapolation of the model's confidence trajectory, and the image-attention mass (Eqs. 8-9) is a model-introspection feature; neither is fit to the reported accuracies. The one genuinely by-construction element is in the internal analysis: the grounding metric m_i is exactly the quantity used to admit safe swap candidates, so the Figure 3 'higher image attention' result is largely entailed by the selection rule rather than being independent evidence of better grounding. This does not transfer to the external benchmark gains. The self-citation to Kim et al. (2026) is motivational and not load-bearing; the method's equations stand alone. The paper's Section A.9 honestly acknowledges that confidence and attention are imperfect proxies, which is a validity limitation rather than a circularity. The cross-table discrepancy for MMaDA MMBench 128/64 (36.55 in Table 1 vs 35.55 in Table 3) and the absence of code/error bars are reproducibility concerns, not circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

ST-Veto introduces no new physical or mechanistic entity. Its free parameters are the benchmark-selected design defaults (attention layers, MAD multiplier, Taylor history length), and the method's correctness rests on two domain assumptions the authors themselves flag as imperfect proxies: confidence-as-correctness and attention-as-grounding.

free parameters (3)
  • Attention layer selection {L/3, L/2, 2L/3} = nearest valid layers at one-third, half, two-thirds of network depth
    Eq. 7-9. The default is justified by the ablation in Table 5, where this combination performs best on most settings of the evaluation benchmarks. It is a design choice selected from the paper's own benchmark results rather than an independently motivated constant.
  • MAD multiplier sigma = 1 for veto/candidate thresholds = 1.0 (default)
    Eq. 11-12. Table 6 compares 0.5MAD, 1.0MAD, 1.5MAD on the evaluation benchmarks and 1.0MAD is chosen as default because it performs best on those benchmarks. The threshold is distribution-adaptive but the multiplier is benchmark-selected.
  • Taylor history length H = 4 (window {t-3, ..., t}) = 4
    Eq. 5. Table 4 shows H=4 is the best cost-performance trade-off on the evaluation benchmarks; H=5 sometimes matches or exceeds it. This is a hand-selected hyperparameter, validated on the test benchmarks.
axioms (4)
  • domain assumption A dMLLM trained for masked-token prediction produces confidence scores that are meaningful and comparable across masked positions and across diffusion steps.
    The entire Veto-and-Swap policy (Eq. 5-14) reads c(t)_i as the model's confidence in the predicted token at position i and treats high confidence as a proxy for correctness. Cited in the method sections 3.1-3.3 and acknowledged as imperfect in Appendix A.9.
  • domain assumption Head-averaged attention mass over image tokens is a valid measure of visual grounding for token selection.
    Eq. 7-9 define the grounding signal exclusively through attention ratios. This imports a large body of prior findings (Favero et al.; Kang et al.; Kim et al.) that attention relates to hallucination, but for dMLLMs this relationship is asserted, not derived. Appendix A.9 notes image attention can be diffuse or semantically misaligned.
  • domain assumption Confidence trajectories of good tokens are smooth enough that a 4-step second-order Taylor prediction is a valid baseline for 'expected' confidence.
    Eq. 5-6 define instability as deviation from a finite-difference Taylor prediction. The paper itself (Appendix A.9) concedes non-smooth confidence trajectories make the Taylor prediction less reliable.
  • domain assumption The base dMLLMs (LaViDa-Reason, MMaDA-MixCoT) are representative of the dMLLM class.
    All empirical conclusions are drawn from exactly two models. The paper's claim that ST-Veto works 'across multiple dMLLMs' is an extrapolation from these two instances (Section 4).

pith-pipeline@v1.3.0-alltime-deepseek · 4472 in / 4540 out tokens · 154601 ms · 2026-08-01T16:43:00.526346+00:00 · methodology

0 comments
read the original abstract

Vision Language Models (VLMs) achieve strong reasoning with Chain-of-Thought (CoT) prompting but incur high sequential-generation cost, error accumulation, and limited self-correction. Diffusion Multimodal Large Language Models (dMLLMs) unmask tokens in an order-agnostic process, improving efficiency and enabling iterative refinement, yet their reasoning and how to enhance it remain underexplored. We propose a training-free method, Spatio-Temporal Token Veto (ST-Veto), which leverages the ability to observe all token positions at each diffusion step. Rather than relying only on current-step confidence, ST-Veto vetoes temporally unstable tokens via second-order Taylor prediction of confidence dynamics and filters weakly grounded tokens using image-attention mass, swapping them with safer candidates. Across multiple dMLLMs and multimodal reasoning benchmarks, ST-Veto consistently outperforms standard decoding policies and prior VLM reasoning methods, improving accuracy by up to 9% with no additional training or generation cost. Analyses show that ST-Veto steers generation toward higher-confidence, better-grounded paths.

Figures

Figures reproduced from arXiv: 2607.17884 by Beomseok Lee, Hyunwoo Kim, Keuntae Kim, Yong Suk Choi.

Figure 1
Figure 1. Figure 1: Overview of our work. The base decoder unmasks tokens according to current-step confidence, whereas ST-Veto applies temporal-stability and visual-grounding checks before committing tokens to the denoising trajectory. 2022) while allowing sufficiently long generation lengths. Crucially, we observe that directly applying state-of-the-art (SOTA) reasoning methods developed for VLMs is ineffec￾tive under the d… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ST-Veto. Tokens are vetoed based on the Taylor residual and low visual grounding, and are replaced by safe tokens that are unmasked instead. However, these methods are structurally incompatible with the diffusion generation paradigm. Unlike AR models that condition current generation on a fixed history of preceding tokens, dMLLMs update all tokens simultaneously through iterative denoising (Sah… view at source ↗
Figure 3
Figure 3. Figure 3: (a) and (b) show the confidence scores of tokens selected at each step for different methods. (c) and (d) present the relative values of the Average Image Attention [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

21 extracted references · 16 linked inside Pith

  1. [4]

    Towards mitigating llm hallucination via self reflection

    Ji, Z., Yu, T., Xu, Y ., Lee, N., Ishii, E., and Fung, P. Towards mitigating llm hallucination via self reflection. InFind- ings of the Association for Computational Linguistics: EMNLP 2023, pp. 1827–1843,

    2023

  2. [5]

    Kang, S., Kim, J., Kim, J., and Hwang, S. J. See what you are told: Visual attention sink in large multimodal models. arXiv preprint arXiv:2503.03321,

    Pith/arXiv arXiv

  3. [8]

    Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., and Gao, J. Llava-med: Training 10 ST-Veto: Spatio-Temporal Token Veto for Diffusion MLLMs via Taylor Prediction and Visual Grounding a large language-and-vision assistant for biomedicine in one day.Advances in Neural Information Processing Systems, 36:28541–28564, 2023a. ...

    Pith/arXiv arXiv

  4. [10]

    Large language diffusion models.arXiv preprint arXiv:2502.09992,

    Nie, S., Zhu, F., You, Z., Zhang, X., Ou, J., Hu, J., Zhou, J., Lin, Y ., Wen, J.-R., and Li, C. Large language diffusion models.arXiv preprint arXiv:2502.09992,

    Pith/arXiv arXiv

  5. [11]

    Pan, L., Saxon, M., Xu, W., Nathani, D., Wang, X., and Wang, W. Y . Automatically correcting large lan- guage models: Surveying the landscape of diverse self- correction strategies.arXiv preprint arXiv:2308.03188,

    Pith/arXiv arXiv

  6. [12]

    Simple and effective masked diffusion language models.Advances in Neural Information Processing Systems, 37:130136– 130184, 2024a

    Sahoo, S., Arriola, M., Schiff, Y ., Gokaslan, A., Marroquin, E., Chiu, J., Rush, A., and Kuleshov, V . Simple and effective masked diffusion language models.Advances in Neural Information Processing Systems, 37:130136– 130184, 2024a. Sahoo, S., Arriola, M., Schiff, Y ., Gokaslan, A., Marroquin, E., Chiu, J., Rush, A., and Kuleshov, V . Simple and effecti...

    arXiv

  7. [13]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191,

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191,

    Pith/arXiv arXiv

  8. [14]

    Self-consistency im- proves chain of thought reasoning in language models

    Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., and Zhou, D. Self-consistency im- proves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171,

    Pith/arXiv arXiv

  9. [16]

    Mmada: Multimodal large diffusion language models.arXiv preprint arXiv:2505.15809,

    Yang, L., Tian, Y ., Li, B., Zhang, X., Shen, K., Tong, Y ., and Wang, M. Mmada: Multimodal large diffusion language models.arXiv preprint arXiv:2505.15809,

    Pith/arXiv arXiv

  10. [17]

    Dream 7b: Diffusion large language models.arXiv preprint arXiv:2508.15487,

    Ye, J., Xie, Z., Zheng, L., Gao, J., Wu, Z., Jiang, X., Li, Z., and Kong, L. Dream 7b: Diffusion large language models.arXiv preprint arXiv:2508.15487,

    Pith/arXiv arXiv

  11. [18]

    You, Z., Nie, S., Zhang, X., Hu, J., Zhou, J., Lu, Z., Wen, J.-R., and Li, C. Llada-v: Large language diffusion 11 ST-Veto: Spatio-Temporal Token Veto for Diffusion MLLMs via Taylor Prediction and Visual Grounding models with visual instruction tuning.arXiv preprint arXiv:2505.16933,

    Pith/arXiv arXiv

  12. [19]

    Vision-language models for vision tasks: A survey.IEEE transactions on pattern analysis and machine intelligence, 46(8):5625– 5644, 2024a

    Zhang, J., Huang, J., Jin, S., and Lu, S. Vision-language models for vision tasks: A survey.IEEE transactions on pattern analysis and machine intelligence, 46(8):5625– 5644, 2024a. Zhang, X., Quan, Y ., Gu, C., Shen, C., Yuan, X., Yan, S., Cheng, H., Wu, K., and Ye, J. Seeing clearly by layer two: Enhancing attention heads to alleviate hallucination in LV...

    Pith/arXiv arXiv

  13. [20]

    Multimodal chain-of-thought reasoning in lan- guage models.arXiv preprint arXiv:2302.00923,

    Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., and Smola, A. Multimodal chain-of-thought reasoning in lan- guage models.arXiv preprint arXiv:2302.00923,

    Pith/arXiv arXiv

  14. [21]

    W→R and R→W denote examples that change from wrong to right and from right to wrong, respectively

    12 ST-Veto: Spatio-Temporal Token Veto for Diffusion MLLMs via Taylor Prediction and Visual Grounding Table 8.Instance-level flip analysis for LaViDa at L/T= 256/128 . W→R and R→W denote examples that change from wrong to right and from right to wrong, respectively. Benchmark Base ST-Veto∆W→R R→W Flip ratio /Z MMBench 31.55 38.25 +6.70 518 228 2.27 / 10.6...

    arXiv 2023

  15. [2000]

    Lei, X., Yang, Z., Chen, X., Li, P., and Liu, Y

    Morgan Kaufmann. Lei, X., Yang, Z., Chen, X., Li, P., and Liu, Y . Scaf- folding coordinates to promote vision-language coor- dination in large multi-modal models.arXiv preprint arXiv:2402.12058,

    Pith/arXiv arXiv

  16. [2021]

    M 3 CoT: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought.arXiv preprint arXiv:2405.16473,

    Chen, Q., Qin, L., Zhang, J., Chen, Z., Xu, X., and Che, W. M 3 CoT: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought.arXiv preprint arXiv:2405.16473,

    Pith/arXiv arXiv

  17. [2022]

    Evaluating object hallucination in large vision-language models

    Li, Y ., Du, Y ., Zhou, K., Wang, J., Zhao, X., and Wen, J.-R. Evaluating object hallucination in large vision-language models. InProceedings of the 2023 Conference on Em- pirical Methods in Natural Language Processing, 2023c. Liu, Y ., Duan, H., Zhang, Y ., Li, B., Zhang, S., Zhao, W., Yuan, Y ., Wang, J., He, C., Liu, Z., et al. Mmbench: Is your multi-m...

    2023

  18. [2023]

    Xu, G., Jin, P., Wu, Z., Li, H., Song, Y ., Sun, L., and Yuan, L

    URL https: //arxiv.org/abs/2312.14135. Xu, G., Jin, P., Wu, Z., Li, H., Song, Y ., Sun, L., and Yuan, L. Llava-cot: Let vision language models reason step- by-step. InProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 2087–2098,

    Pith/arXiv arXiv 2087

  19. [2024]

    MME: A comprehensive evaluation benchmark for multimodal large language models.arXiv preprint arXiv:2306.13394,

    Fu, C., Chen, P., Shen, Y ., Qin, Y ., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., and Ji, R. MME: A comprehensive evaluation benchmark for multimodal large language models.arXiv preprint arXiv:2306.13394,

    Pith/arXiv arXiv

  20. [2025]

    Diffuseq: Sequence to sequence text generation with diffusion mod- els.arXiv preprint arXiv:2210.08933,

    Gong, S., Li, M., Feng, J., Wu, Z., and Kong, L. Diffuseq: Sequence to sequence text generation with diffusion mod- els.arXiv preprint arXiv:2210.08933,

    Pith/arXiv arXiv

  21. [2026]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In Langley, P. (ed.),Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp. 1207–1216, Stan- ford, CA,

    2000