Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Errors on counterfactual images are pulled toward the model's no-image answer, and selectively subtracting that estimated prior repairs them while sparing commonsense 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

2026-08-03 10:58 UTC pith:V3IRJE65

load-bearing objection A solid, honestly-scoped method for counterfactual-image MCQ correction; the QA claims are over-inflated and the prior proxy is fragile, but the MC evidence and self-critique earn it a real referee. the 4 major comments →

arxiv 2607.29240 v1 pith:V3IRJE65 submitted 2026-07-31 cs.CV cs.AI

When Model Priors Conflict with Visual Evidence: Mitigating Commonsense-Driven Hallucinations by Selective Prior Calibration

Kesheng Chen , Yamin Hu , Wenjian Luo This is my paper
classification cs.CV cs.AI
keywords commonsense-driven hallucinationvision-language modelscounterfactual imagesprior calibrationhallucination mitigationdirected prior attractionselective prior calibrationvisual priors
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.

Commonsense-driven hallucination is when a vision-language model overrides clear visual evidence of an atypical state and reports the ordinary state instead — a visibly six-fingered hand is reported as five-fingered. The paper shows these errors are directed: when a model gets a counterfactual image wrong, its answer usually coincides with the candidate it scores highest with the image removed. It then introduces Selective Prior Calibration (SPC), which subtracts these text-only prior-preference scores from image-conditioned candidate scores with a learned, instance-dependent strength and revises the prediction only when the corrected score pattern strongly supports an alternative. On a benchmark of 300 matched counterfactual/commonsense image pairs, SPC raises counterfactual accuracy by 7.8 and 7.0 points on two large models while changing commonsense accuracy by only +0.4 and −0.4 points. If the reported error direction is a general property of vision-language models, this offers a cheap, training-free way to reduce a specific class of hallucinations without retraining.

Core claim

The central claim is directed prior attraction: incorrect predictions on counterfactual images systematically converge on the candidate preferred by the model's text-only scoring, and this candidate usually equals the ground-truth answer of the matched commonsense image. Suppressing this prior indiscriminately trades one error for another, but suppressing it with an instance-dependent strength — learned from paired counterfactual/commonsense development data — and gating the change behind a support condition repairs counterfactual errors while preserving commonsense accuracy. On the benchmark's multiple-choice task, corrected counterfactual accuracy rises by 7.8 points for Qwen3-VL-32B and 7

What carries the argument

The load-bearing identity is sC(y) = s̄I(y) − λ s̄P(y): the corrected score of candidate y equals its mean-centered image-conditioned log-probability minus λ times its mean-centered text-only log-probability, the latter being the operational estimate of the model's prior preference. λ is a per-instance correction strength produced by a low-capacity softplus-linear model over 13 score-derived features, fit by maximum-a-posteriori estimation with a Gaussian prior on paired CF/CS development data. Around that sits a selective revision rule: the corrected proposal is adopted only if it differs from the original answer, its softmax margin over the runner-up exceeds a threshold, the candidate-set

Load-bearing premise

That the model's score for a candidate with the image removed (text-only log-probability) faithfully measures the latent commonsense prior that pulls errors toward the ordinary-state answer, even when question wording or claim order changes.

What would settle it

Build a set of counterfactual images where the text-only scores do not favor the ordinary-state answer but the model still errs toward the ordinary answer; if SPC's correction then fails on those cases, directed prior attraction is not the error mechanism. The paper's own RQ5 approaches this condition for LLaVA under 'Instead,' 'Choice,' and 'Reject second' question forms, where the prior favors the ordinary answer in only 14.3%, 5.7%, and 0.0% of pairs and CF gains disappear or go negative on atypical-first questions.

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

If this is right

  • CDH errors are not random: on the benchmark, 70–92% of incorrect counterfactual answers select the text-only prior-favored candidate across models and tasks.
  • Candidate-level prior subtraction, not just re-ranking by image-conditioned scores, drives the gain; ablated variants without the subtraction lose or reverse the improvement.
  • Instance-dependent strength and gated revision are complementary: always adopting the corrected proposal repairs more counterfactual errors but harms matched commonsense images (e.g., LLaVA QA commonsense loss jumps to −13.9 points).
  • The correction transfers to held-out CDH categories, unseen candidate-order permutations, and other conflict benchmarks, while changing almost no predictions on object-hallucination benchmarks.
  • The text-only prior estimate is the fragile link: under rephrased questions and reversed claim orders, correction gains shrink or reverse, and for one model the text-only prediction reverses with claim order on only 0–47% of cases depending on question form.

Where Pith is reading between the lines

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

  • If directed prior attraction is a general failure mode, SPC suggests a diagnostic: comparing a vision-language model's image-conditioned and text-only answer distributions on any QA pair quantifies how much a given question leans on priors, independent of benchmark labels.
  • Using multiple rephrasings of the same question to estimate the prior (an ensemble over forms) might stabilize the correction where single-form text-only scores mislead — a testable extension of the paper's own RQ5 analysis.
  • Because SPC needs only candidate log-probabilities, it could be combined with decoding-time interventions rather than only post-hoc candidate rescoring; the paper's ablations show that hand-designed token-level contrastive rules change far more predictions on object-hallucination benchmarks, suggesting a calibration target for such methods.
  • If text-only score order reliably tracked the ordinary-state answer, SPC would amount to an unsupervised visual-fidelity check; the RQ5 results bound when that proxy is safe, for instance favoring 'contrastive' question forms over rephrased alternatives.

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 / 4 minor

Summary. The paper studies commonsense-driven hallucination (CDH) in vision-language models, where a model answers according to its text-only commonsense prior even when the image supports an atypical state. It proposes Selective Prior Calibration (SPC): for each candidate answer, compute image-conditioned and text-only log-probabilities, mean-center them, subtract a scaled version of the text-only scores from the image-conditioned scores with an instance-dependent strength learned on paired CF-CS development data, and adopt the revised proposal only if it satisfies margin and support conditions. Experiments on CDH-Bench MC/QA, Visual CounterFact, HallusionBench, ConflictVIS, POPE, and POPEv2 use Qwen3-VL-32B and LLaVA-1.6-34B, with ablations, bootstrap intervals, and McNemar tests. The central claims are directed prior attraction, effective CF repair with CS retention, and generalization across categories, candidate-order permutations, and external conflict benchmarks.

Significance. If the result holds, SPC provides a simple, training-free-at-inference method with a clear mechanism and a favorable repair-retention trade-off. The paper's strengths are substantial: code release, exact McNemar tests and bootstrap intervals, a careful component analysis isolating candidate-specific prior subtraction, adaptive strength, and selective revision, plus transfer to external conflict benchmarks. The paper is also unusually honest in RQ5, showing where the method fails. However, the result is conditional on the text-only prior proxy remaining aligned with the ordinary-state preference, and the main text overstates the QA and generalization claims relative to the evidence in the supplementary material.

major comments (4)
  1. [Table 2 and Appendix E] The NoLan baseline in Table 2 is not measured against the same reference as the other entries. Table 2 reports NoLan as +4.8/-1.7 (Qwen MC), +15.2/-0.9 (Qwen QA), +11.7/-3.5 (LLaVA MC), and +24.8/-8.7 (LLaVA QA). Appendix E states that these values correspond to a matched decoding loop with beta=0, while the changes relative to the common original predictions are +0.0/-4.8, +13.9/-0.9, +9.6/-2.6, and +24.8/-8.7. Since all other rows are measured against the original model, the 'SPC achieves the highest utility' claim is not based on a like-for-like comparison. The authors should report one consistent reference for all methods or justify why a different reference is appropriate for NoLan.
  2. [Appendix H, Table A17/A18; Section 5 QA rows] The binary-QA headline numbers are confounded by label imbalance. In the original QA form, 218 of 230 CF examples have ground truth 'no' and only 12 have 'yes.' SPC changes Qwen accuracy by +26.1 on the no side and -16.7 on the yes side; LLaVA changes by +32.6 and -16.7. The balanced claim-order test shows that the LLaVA atypical-first gain is not reliable: after fitting on the original form, SPC changes atypical-first CF by -1.7 (95% interval [-6.5, 3.0]). The paper acknowledges this only in the appendix and in a sentence in Section 6. The main text should not present the QA rows of Tables 1-2 as general evidence of CF repair; the balanced test should be reported in the main body, and the abstract/conclusion should be qualified accordingly.
  3. [Section 5, Table 6 / Table A20; Eq. (1)] The central mechanism requires the text-only log-probability s_P(y) to be a stable estimate of the model's latent preference for the ordinary state. RQ5 shows this assumption breaks under question-form and claim-order changes. For LLaVA, the text-only prediction reverses with claim order on only 47.4% of Contrastive pairs, and on 14.3%, 5.7%, and 0.0% of the three unseen forms; SPC's CF accuracy change on atypical-first questions is -11.3, -13.9, and -43.0 points, respectively, with large inconsistent CS changes. The abstract's unqualified claim that SPC 'substantially improves accuracy on CF images' is therefore too broad. The conclusion should state explicitly that the method is reliable when the prior proxy is aligned with the ordinary-state answer, and the atypical-first results should be summarized in the main text rather than only in the supplement.
  4. [Section 3, Eq. (5); RQ1] The 'directed prior attraction' finding is close to definitional and should be framed more carefully. CDH errors are, by construction, cases in which the commonsense prior overrides visual evidence, and the text-only score is the same model's own commonsense answer preference. The empirical content is not the high attraction rate per se, but rather that the prior-favored candidate coincides with the paired CS ground truth; the paper does show this in Figure 2 and the appendix. Given that CDH-Bench is from the same research group (Chen et al. 2026), the discovery claim should be tempered and the external benchmarks emphasized as the stronger evidence for the mechanism.
minor comments (4)
  1. [Section 4 / Table 1] CFAD, RPD, and CCR are used in the main text but defined only in the appendix (Table A2 caption). Define them at first use in the main paper.
  2. [Figure 2] Several cells have very small n (e.g., n=1 or n=2). The text notes this for attribute cells, but the figure would be clearer if cells with n<3 were visually de-emphasized or marked.
  3. [Appendix C] The uncertainty-aware extension is used for generalization, category-holdout, and reduced-data results, but the main paper does not always state this. A sentence in Section 4 clarifying which results use the point-estimate SPC and which use the uncertainty-aware variant would improve reproducibility.
  4. [Section 5, RQ4 and Table 5] The notation 'Reported change' in Table 5 mixes CF/CS deltas and overall accuracy deltas without a uniform unit. This is understandable per benchmark, but the caption should state explicitly which metric each row uses.

Circularity Check

0 steps flagged

No significant circularity: the CF gains are held-out test results, the directed-prior-attraction measurement is an empirical correlation between independently defined quantities, and the CDH-Bench self-citation is corroborated by external Visual CounterFact.

full rationale

The paper's derivation chain is: (1) define CF errors as incorrect answers on counterfactual images, independently of any text-only score; (2) measure how often those errors coincide with the text-only top candidate (RQ1); (3) propose a correction that subtracts text-only scores with a learned strength (Eq. 1); (4) fit λ and the revision threshold on a disjoint 70-pair development set; (5) evaluate on the held-out 230 test pairs and on external benchmarks. No equation or fitted parameter is renamed as a prediction. The success of SPC is not tautological: subtracting the prior score from the prior-favored candidate does not by itself guarantee repair, because the corrected alternative must also have sufficient image-conditioned support and pass the selective revision gate; the ablations show that candidate-specific subtraction matters and that a fixed or ungated correction can hurt CS accuracy. The paper's RQ5 finding that the text-only proxy is unstable across question forms and claim orders is a scope and robustness limitation, not a circular step: it narrows the conditions under which the method works but does not show that the test-set CF gains reduce to the training objective by construction. The CDH-Bench is cited from overlapping authors, but the central 'directed prior attraction' claim is independently corroborated on Visual CounterFact, where all 121 original CF errors select the ordinary attribute, so the self-citation is not load-bearing. Overall, the paper's claimed predictions are evaluated on held-out data and external benchmarks, and no specific reduction of a result to its own inputs is exhibited.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

SPC introduces no new physical entities; its load-bearing assumptions are the operational proxy for the prior and the correctness of the self-built CDH-Bench labels. The free parameters are the learned correction strength and the dev-selected gates (λmax, mmin, ρ, ϵCS).

free parameters (5)
  • θ (13-dim correction-strength weights) = learned via MAP on 70 CDH-Bench dev pairs; exact values not printed
    λθ(u)=softplus(u^T θ) determines how much prior to subtract; central to Eq. (1)-(4).
  • λmax (max correction strength) = 1.0 (Qwen), 0.8 (LLaVA) at ϵCS=0.04
    Caps the learned strength; selected on dev (Table A1).
  • mmin (minimum proposal margin) = 0.3 (Qwen), 0.2 (LLaVA)
    Change gate in Eq. (8); selected on dev.
  • ρ (regularization) = 0.01 both models
    Gaussian prior strength for θ; selected on dev.
  • ϵCS (allowed CS loss) = 0.04 primary, 0.10 secondary
    Selection constraint; authors report both; affects reported numbers.
axioms (5)
  • domain assumption CDH-Bench CF/CS ground-truth answers are visually correct
    All headline CF/CS accuracy numbers treat this benchmark as ground truth; benchmark was constructed by the same group (Chen et al. 2026).
  • domain assumption Text-only candidate log-probabilities (image removed) estimate the model's latent prior
    SPC's correction and the directed-prior-attraction measurement both rely on this proxy; Appendix B Eq. A1 writes s̄P=p.
  • standard math Mean-centering and softmax over candidates preserves decision-relevant score information
    Used to define sC and qC; candidate rankings unchanged by centering, but absolute score calibration across examples is assumed.
  • domain assumption Greedy decoding with temperature 0 and parsing the first valid token gives the model's intended answer
    Original predictions and baselines are parsed this way; unparseable outputs are counted wrong (Appendix A).
  • domain assumption Matched CF-CS pairs share all factors except the depicted state
    The CS-retention metric assumes the CS image is the correct control for the CF image.

pith-pipeline@v1.3.0-daily-deepseek · 36564 in / 13653 out tokens · 127281 ms · 2026-08-03T10:58:17.128951+00:00 · methodology

0 comments
read the original abstract

In vision--language models, commonsense-driven hallucination (CDH) occurs when a model's commonsense prior overrides clear visual evidence of an atypical state. For example, a model may report that a visibly six-fingered hand has five fingers. We show that these errors are systematically directed: when a model answers a question about a counterfactual (CF) image incorrectly, its answer often coincides with the candidate it prefers without access to the image. Suppressing this prior indiscriminately can repair CF errors, but may also disrupt correct answers on matched commonsense (CS) images, where the same prior is helpful. We therefore propose Selective Prior Calibration (SPC), which subtracts candidate-level prior-preference estimates from image-conditioned scores with an instance-dependent strength and revises the original prediction only when the resulting score pattern strongly supports an alternative. Extensive experiments demonstrate that SPC substantially improves accuracy on CF images while largely preserving accuracy on matched CS images. Furthermore, these gains generalize across CDH categories, candidate-answer permutations, and other conflict benchmarks, while SPC rarely alters predictions on benchmarks without such conflicts.

Figures

Figures reproduced from arXiv: 2607.29240 by Kesheng Chen, Wenjian Luo, Yamin Hu.

Figure 1
Figure 1. Figure 1: From directed prior attraction to selective prior calibration. (a) On the CS image, visual evidence aligns with the model’s prior. On the CF image, they conflict, and the same prior biases the model toward the prior-favored answer. (b) SPC scores candidates with and without the image, adjusts their scores using a learned correction strength, and adopts the corrected proposal only when it is sufficiently su… view at source ↗
Figure 2
Figure 2. Figure 2: Prior attraction by CDH subcategory. Each cell reports the fraction of original CF errors that select the no￾image prior winner. Rows represent CDH subcategories, and columns represent model–task settings. +0.4 and -0.4 points. Repairs substantially outnumber harms (21/3 and 17/1). Consistent with these gains, CFAD and RPD decrease, indicating a narrower CF–CS performance gap, while the lower CCR shows tha… view at source ↗
Figure 3
Figure 3. Figure 3: Fitting with CF instances alone increases repair [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of prior-direction controls on Qwen3-VL￾32B under MC.Box plots show the distributions over 10,000 pair-level bootstrap resamples. marks, we select all configurations on CDH-Bench and apply them to other benchmarks without retuning. Generalization to held-out CDH categories [PITH_FULL_IMAGE:figures/full_fig_p006_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

39 extracted references · 5 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2603.27982 , year =

    Chen, Kesheng and Hu, Yamin and Zhou, Qi and Zhu, Zhenqian and Luo, Wenjian , title =. arXiv preprint arXiv:2603.27982 , year =. doi:10.48550/arXiv.2603.27982 , url =. 2603.27982 , archivePrefix =

    doi:10.48550/arxiv.2603.27982

  2. [2]

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages =

    Rohrbach, Anna and Hendricks, Lisa Anne and Burns, Kaylee and Darrell, Trevor and Saenko, Kate , title =. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages =. 2018 , publisher =

    2018

  3. [3]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages =

    Li, Yifan and Du, Yifan and Zhou, Kun and Wang, Jinpeng and Zhao, Wayne Xin and Wen, Ji-Rong , title =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages =. 2023 , publisher =

    2023

  4. [4]

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

    Guan, Tianrui and Liu, Fuxiao and Wu, Xiyang and Xian, Ruiqi and Li, Zongxia and Liu, Xiaoyu and Wang, Xijun and Chen, Lichang and Huang, Furong and Yacoob, Yaser and Manocha, Dinesh and Zhou, Tianyi , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  5. [5]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

    Agrawal, Aishwarya and Batra, Dhruv and Parikh, Devi and Kembhavi, Aniruddha , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

  6. [6]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

    Goyal, Yash and Khot, Tejas and Summers-Stay, Douglas and Batra, Dhruv and Parikh, Devi , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =. 2017 , doi =

    2017

  7. [7]

    Advances in Neural Information Processing Systems , volume =

    Cad. Advances in Neural Information Processing Systems , volume =

  8. [8]

    and Lee, Stefan and Shen, Yilin and Jin, Hongxia and Ghosh, Shalini and Heck, Larry and Batra, Dhruv and Parikh, Devi , title =

    Selvaraju, Ramprasaath R. and Lee, Stefan and Shen, Yilin and Jin, Hongxia and Ghosh, Shalini and Heck, Larry and Batra, Dhruv and Parikh, Devi , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  9. [9]

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

    Chen, Long and Yan, Xin and Xiao, Jun and Zhang, Hanwang and Pu, Shiliang and Zhuang, Yueting , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  10. [10]

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

    Niu, Yulei and Tang, Kaihua and Zhang, Hanwang and Lu, Zhiwu and Hua, Xian-Sheng and Wen, Ji-Rong , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  11. [11]

    Shortcut Learning in Deep Neural Networks , journal =

    Geirhos, Robert and Jacobsen, J. Shortcut Learning in Deep Neural Networks , journal =. 2020 , doi =

    2020

  12. [12]

    arXiv preprint arXiv:2311.07397 , year =

    Wang, Junyang and Wang, Yuhang and Xu, Guohai and Zhang, Jing and Gu, Yukai and Jia, Haitao and Wang, Jiaqi and Xu, Haiyang and Yan, Ming and Zhang, Ji and Sang, Jitao , title =. arXiv preprint arXiv:2311.07397 , year =

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2306.13394 , year =

    Fu, Chaoyou and Chen, Peixian and Shen, Yunhang and Qin, Yulei and Zhang, Mengdan and Lin, Xu and Yang, Jinrui and Zheng, Xiawu and Li, Ke and Sun, Xing and Wu, Yunsheng and Ji, Rongrong , title =. arXiv preprint arXiv:2306.13394 , year =

    Pith/arXiv arXiv

  14. [14]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages =

    Sun, Zhiqing and Shen, Sheng and Cao, Shengcao and Liu, Haotian and Li, Chunyuan and Shen, Yikang and Gan, Chuang and Gui, Liangyan and Wang, Yu-Xiong and Yang, Yiming and Keutzer, Kurt and Darrell, Trevor , title =. Findings of the Association for Computational Linguistics: ACL 2024 , pages =. 2024 , publisher =

    2024

  15. [15]

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

    Leng, Sicong and Zhang, Hang and Chen, Guanzheng and Li, Xin and Lu, Shijian and Miao, Chunyan and Bing, Lidong , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  16. [16]

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

    Favero, Alessandro and Zancato, Luca and Trager, Matthew and Choudhary, Siddharth and Perera, Pramuditha and Achille, Alessandro and Swaminathan, Ashwin and Soatto, Stefano , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , doi =

    2024

  17. [17]

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

    Huang, Qidong and Dong, Xiaoyi and Zhang, Pan and Wang, Bin and He, Conghui and Wang, Jiaqi and Lin, Dahua and Zhang, Weiming and Yu, Nenghai , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  18. [18]

    Proceedings of the 41st International Conference on Machine Learning , series =

    Chen, Zhaorun and Zhao, Zhuokai and Luo, Hongyin and Yao, Huaxiu and Li, Bo and Zhou, Jiawei , title =. Proceedings of the 41st International Conference on Machine Learning , series =. 2024 , publisher =

    2024

  19. [19]

    Computer Vision -- ECCV 2024 , pages =

    Liu, Shi and Zheng, Kecheng and Chen, Wei , title =. Computer Vision -- ECCV 2024 , pages =. 2024 , publisher =

    2024

  20. [20]

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

    Liu, Bingqian and Zhang, Fu and Chen, Guoqing and Cheng, Jingwei , title =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages =. 2025 , publisher =

    2025

  21. [21]

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

    Yang, Le and Zheng, Ziwei and Chen, Boxu and Zhao, Zhengyu and Lin, Chenhao and Shen, Chao , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  22. [22]

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

    Chen, Junzhe and Zhang, Tianshu and Huang, Shiyu and Niu, Yuwei and Zhang, Linfeng and Wen, Lijie and Hu, Xuming , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  23. [23]

    arXiv preprint arXiv:2602.11824 , year =

    Wu, Jialin and Shi, Wei and Shen, Han and Qi, Peigui and Tang, Kunsheng and Huang, Zhicong and Wang, Binghao and Yang, Zhou , title =. arXiv preprint arXiv:2602.11824 , year =

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2602.22144 , year =

    Ren, Lingfeng and Yu, Weihao and Yu, Runpeng and Wang, Xinchao , title =. arXiv preprint arXiv:2602.22144 , year =. doi:10.48550/arXiv.2602.22144 , url =. 2602.22144 , archivePrefix =

    doi:10.48550/arxiv.2602.22144

  25. [25]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages =

    Zhou, Kankan and Lai, Eason and Yeong, Wei Bin Au and Mouratidis, Kyriakos and Jiang, Jing , title =. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages =. 2023 , publisher =

    2023

  26. [26]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages =

    Lee, Kang-il and Kim, Minbeom and Yoon, Seunghyun and Kim, Minsung and Lee, Dongryeol and Koh, Hyukhun and Jung, Kyomin , title =. Findings of the Association for Computational Linguistics: NAACL 2025 , pages =. 2025 , publisher =

    2025

  27. [27]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics , pages =

    Liu, Xiaoyuan and Wang, Wenxuan and Yuan, Youliang and Huang, Jen-tse and Liu, Qiuzhi and He, Pinjia and Tu, Zhaopeng , title =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics , pages =. 2025 , publisher =

    2025

  28. [28]

    The Thirteenth International Conference on Learning Representations , year =

    Zhou, Guanyu and Yan, Yibo and Zou, Xin and Wang, Kun and Liu, Aiwei and Hu, Xuming , title =. The Thirteenth International Conference on Learning Representations , year =

  29. [29]

    arXiv preprint arXiv:2605.29579 , year =

    Zhou, Shizhe and Jia, Bohan and Wu, Kai and Shen, Yan and Li, Tongyun and Wu, Yuyang and Lin, Shaohui , title =. arXiv preprint arXiv:2605.29579 , year =

    Pith/arXiv arXiv

  30. [30]

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

    Golovanevsky, Michal and Rudman, William and Lepori, Michael and Bar, Amir and Singh, Ritambhara and Eickhoff, Carsten , title =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages =. 2025 , publisher =

    2025

  31. [31]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages =

    Zheng, Kening and Chen, Junkai and Yan, Yibo and Zou, Xin and Zhou, Huiyu and Hu, Xuming , title =. Findings of the Association for Computational Linguistics: ACL 2025 , pages =. 2025 , publisher =

    2025

  32. [32]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =

    Liu, Shuliang and Yang, Songbo and Fang, Dong and Jia, Sihang and Tang, Yuqi and Su, Lingfeng and Peng, Ruoshui and Yan, Yibo and Zou, Xin and Hu, Xuming , title =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =. 2026 , publisher =

    2026

  33. [33]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =

    Ortu, Francesco and Jin, Zhijing and Doimo, Diego and Cazzaniga, Alberto , title =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =. 2026 , doi =

    2026

  34. [34]

    Arbitration Failure, Not Perceptual Blindness: How Vision-Language Models Resolve Visual-Linguistic Conflicts , journal =

    Nooralahzadeh, Farhad and Rohanian, Omid and Zhang, Yi and F. Arbitration Failure, Not Perceptual Blindness: How Vision-Language Models Resolve Visual-Linguistic Conflicts , journal =. 2026 , url =

    2026

  35. [35]

    arXiv preprint arXiv:2511.21631 , year =

    Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and Chen, Keqin and Chen, Xionghui and Cheng, Zesen and Deng, Lianghao and Ding, Wei and Gao, Chang and Ge, Chunjiang and others , title =. arXiv preprint arXiv:2511.21631 , year =

    Pith/arXiv arXiv

  36. [36]

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

    Liu, Haotian and Li, Chunyuan and Li, Yuheng and Lee, Yong Jae , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  37. [37]

    Transformer Circuits Thread , year =

    Gurnee, Wes and Sofroniew, Nicholas and Pearce, Adam and Piotrowski, Mateusz and Kauvar, Isaac and Chen, Runjin and Soligo, Anna and Bogdan, Paul and Ong, Euan and Wang, Rowan and Thompson, Ben and Abrahams, David and Kantamneni, Subhash and Ameisen, Emmanuel and Batson, Joshua and Lindsey, Jack , title =. Transformer Circuits Thread , year =

  38. [38]

    Findings of the Association for Computational Linguistics: ACL 2026 , pages =

    Fazli, Mehrdad and Wei, Bowen and Zhu, Ziwei , title =. Findings of the Association for Computational Linguistics: ACL 2026 , pages =. 2026 , publisher =

    2026

  39. [39]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Li, Yifan and Zhou, Kun and Zhao, Wayne Xin and Fang, Lei and Wen, Ji-Rong , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2026 , doi =

    2026