Pith. sign in

REVIEW 5 major objections 5 minor 80 references

Multimodal AI models still fail to balance answering and refusal when the image context shifts.

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

2026-08-04 01:18 UTC pith:SNFTFBSL

load-bearing objection MMOOC is a useful benchmark with a genuine taxonomy; the central claim survives, but missing label-reliability statistics need to be addressed before the fine-grained numbers can be trusted. the 5 major comments →

arxiv 2607.27637 v2 pith:SNFTFBSL submitted 2026-07-30 cs.CV

MMOOC: A Comprehensive Benchmark for Out-of-Context Evaluation in Multimodal Large Language Models

Wenjie Zhu , Yabin Zhang , Wenjun Zeng , Lei Zhang This is my paper
classification cs.CV
keywords multimodal large language modelsout-of-context evaluationrefusalanswerabilitybenchmarkhallucinationshifted in-contextvisual question answering
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 a reliable multimodal model must do two things at once: refuse genuinely out-of-context questions, where the image cannot support an answer, and still answer shifted in-context questions, where a distracting false premise or mismatch is present but the core visual fact remains observable. To measure this balance, the authors built MMOOC, a 41K image-question benchmark spanning three question formats, eight shift types, and six visual scenarios, with five out-of-context categories and three shifted in-context categories. Across 18 current models, they find low and inconsistent refusal rates, especially on uncertain spatial and physical context and unclear logical and symbolic cases, and limited accuracy on partially answerable questions. The paper's central claim is that current MLLMs still struggle to balance robust answering and appropriate refusal, and MMOOC provides a reusable instrument to quantify that struggle.

Core claim

MMOOC establishes that current multimodal large language models systematically mishandle shifted contexts: they refuse too rarely when questions are truly unanswerable, and they answer too poorly when a question remains answerable despite distracting or misleading context. The benchmark defines three answerable shifted in-context categories (misleading premise, partial answerability, image-question mismatch) and five unanswerable out-of-context categories (multimodal ambiguity, visual false premises, uncertain spatial and physical context, unclear logical and symbolic, missing knowledge and background). Evaluation uses Accuracy and Answer Rationality for in-context samples and Refusal Rate a

What carries the argument

The central object is MMOOC's hierarchical taxonomy of answerability under context shift. Samples are first classified as shifted in-context if the answer remains supported by image and question evidence, or out-of-context if the available evidence is insufficient even with general world knowledge; each sample then receives one of eight fine-grained shift types. This taxonomy is paired with a dual-metric evaluation protocol—Accuracy/Answer Rationality for answerable cases and Refusal Rate/Refusal Rationality for unanswerable ones—so a model is scored on both knowing when to refuse and knowing when to answer. Data quality is enforced by retaining samples only when three large models independe

Load-bearing premise

The ground-truth answerability labels and reference answers are correct: samples were kept only when three models agreed and then were human-verified, but no inter-annotator agreement is reported, so systematically wrong labels would invalidate the measured refusal rates and accuracies.

What would settle it

Independently re-annotate a random sample of roughly 500 MMOOC items with human annotators who do not see the original labels; if their answerability judgments disagree with the dataset labels on more than a small percentage of items, the benchmark's core metrics lose their meaning. Alternatively, if a perception-weak model achieves high refusal scores on question-only input, that would indicate the metric is partly driven by textual cues rather than true context evaluation.

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

If this is right

  • Current MLLMs show low and inconsistent refusal rates on out-of-context questions, with the most acute failures in uncertain spatial and physical context and unclear logical and symbolic cases.
  • Shifted in-context accuracy, particularly partial answerability in open-ended VQA, remains limited, so contextual shifts disrupt answering even when sufficient visual evidence exists.
  • Stronger general reasoning ability and larger model scale do not consistently translate into better out-of-context robustness; proprietary models can underperform open-weight models on specific categories.
  • Supervised fine-tuning and preference optimization improve refusal behavior, but can trade off against general multimodal capability; CoT prompting offers a better answer-refusal balance than explicit refusal instructions.
  • MMOOC provides a reusable evaluation for calibrating answerability versus refusal, and its question-only baseline suggests some models may rely on textual cues rather than genuine multimodal evidence evaluation.

Where Pith is reading between the lines

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

  • If MMOOC's answerability labels are sound, then standard leaderboards that report only accuracy on answerable questions are missing a central reliability dimension; future evaluation should jointly report refusal and accuracy under shifted contexts.
  • The finding that some models refuse at high rates with no image input hints that refusal can become a superficial textual heuristic; a natural test is to measure the gap between image-present and image-absent refusal to isolate genuine visual grounding.
  • The observed SFT-versus-general-capability trade-off suggests a 'refusal-utility' frontier; a testable extension is to construct preference pairs that reward answering supported sub-queries while refusing unsupported ones, potentially pushing both metrics up simultaneously.
  • The taxonomy transfers naturally to video, audio, and embodied settings, where spatial/physical uncertainty and partial answerability are even more salient; extending MMOOC to temporal modalities would test whether the same failure patterns persist.

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

5 major / 5 minor

Summary. The paper introduces MMOOC, a 41K-scale benchmark for evaluating multimodal large language models under out-of-context (OOC) and shifted in-context (IC) conditions. It defines five OOC categories and three shifted IC categories across three question formats and six visual scenarios, with data generated by multiple MLLMs, filtered by unanimous answerability judgments from GPT-4o/o1/o3, and then verified by human annotators. The evaluation uses Accuracy and Answer Rationality for IC samples and Refusal Rate and Refusal Rationality for OOC samples, with an LLM-as-a-Judge protocol validated on a 2,000-response human sample. Experiments on 18 open-source and closed-source MLLMs report that current models struggle to balance answering and refusal, that OOC performance is highly sensitive to format and category, that stronger general capability does not reliably imply OOC robustness, and that SFT/DPO and prompting can partially improve behavior.

Significance. If the binary answerability labels and reference answers are reliable, MMOOC fills a real gap: existing refusal benchmarks mostly target unanswerable questions, while MMOOC jointly measures refusal on truly unanswerable OOC items and robust answering on answerable but perturbed IC items. The benchmark is large, spans a sensible taxonomy, and the paper includes several good practices: multiple judges that are disjoint from the evaluated model set, a human agreement check, a question-only diagnostic baseline, and post-training comparisons. However, the central quantitative claims rest on label quality and on per-cell statistics that are not currently reported. The paper's headline conclusion—that MLLMs still struggle to balance robust answering and appropriate refusal—is plausible but not yet fully evidenced without stronger label-reliability analysis, per-category sample sizes, and control for linguistic shortcutting.

major comments (5)
  1. [§3.3, Eqs. (1) and (4)] The binary OOC/IC answerability label is the load-bearing assumption: both Accuracy and Refusal Rate compare model behavior against this label. The manuscript reports that samples are retained only when GPT-4o, o1, and o3 agree, and then human annotators verify, but it gives no inter-annotator agreement, no number of annotators, no counts of removed or corrected samples, and no error analysis for the final labels. Because GPT-4o/o1/o3 also generated the data and are among the evaluated models, the potential for label contamination is non-trivial. Please report human agreement (e.g., Cohen's kappa), a breakdown of filtration decisions, and a post-hoc audit of a random sample of final labels. Without this, low OOC refusal rates and low shifted-IC accuracies could partly reflect label noise rather than model failure.
  2. [Table A1 and §4.3] The question-only analysis is a good diagnostic, but as reported it is too coarse to support the paper's visual-context interpretation. Several models show high question-only refusal (e.g., Gemma-4-26B at 92.00, o1 at 78.00), suggesting that many OOC items can be recognized from linguistic cues alone. The table only reports aggregate refusal per model, with no per-category breakdown and no comparison against the image-condition scores. Please provide per-category question-only results and a delta analysis (e.g., image-minus-question-only) so that the reader can see whether the reported OOC category differences reflect genuine visual-context evaluation or linguistic priors.
  3. [Tables 3–4, A2–A7] No per-cell sample sizes, confidence intervals, or significance tests are reported anywhere. Many numbers in the appendix are exact multiples of 2, suggesting per-cell sample sizes of around 50; under that assumption, differences of a few points are within binomial noise. For example, Qwen3-VL-8B versus Qwen3-VL-30B on USPC-YN (15.75 vs. 22.50) may not be meaningful without error bars. The claims that OOC performance is inconsistent, that model scale does not help, and that VQA is easier than YesNo are all point-estimate comparisons. Please add per-cell N, confidence intervals, and, where possible, significance tests or effect sizes.
  4. [§3.5, Eq. (1)] Accuracy is defined as an exact indicator match, but no matching or normalization protocol is given for open-ended VQA answers. Exact string matching is known to underestimate accuracy for free-form answers unless the benchmark defines synonym handling, numerical tolerance, and abbreviation normalization. Since Accuracy contributes half of the IC score S_IC, an unspecified matching rule makes the shifted-IC results hard to interpret. Please specify the exact-match criterion or replace it with a judge-based correctness check.
  5. [§4.3, Table 5] The post-training experiments are presented as evidence that SFT/DPO improve OOC robustness, but the table reports only three base models with no sample sizes, confidence intervals, or significance testing. Given the small per-cell counts implied by the appendix tables, the differences (e.g., Qwen3-VL-2B Rref 8.00 to 38.00 under SFT) need uncertainty quantification. Please report the number of training and evaluation samples and add confidence intervals or a significance test.
minor comments (5)
  1. [§3.3] The human verification procedure is described in one sentence. Even if full agreement statistics are added for the response-judge validation, please also describe the annotation instructions, the number of annotators, and how borderline cases were adjudicated.
  2. [Table A8] The table header appears malformed: 'YesNo MCQ VQA' is printed as if it were a row label, while the columns are the three mismatch sources. Please restructure the table so that the question format is a column or a clear grouping.
  3. [References] Reference [20] is a different paper also titled 'MMOOC' (a multimodal misinformation dataset). The naming collision should be acknowledged in the text or the benchmark should be given a more distinctive name to avoid confusion.
  4. [Limitations] The Limitations subsection only mentions extending to video/audio/embodied environments. It should also acknowledge the current dependence on LLM-generated/filtered labels and the question-only shortcut issue, and point to the supplemental analyses that address them.
  5. [Figure 3 / Table 2] The arrows in 'Dataset Length↑', 'Self-BLEU↓', and 'Similarity↓' are not defined in the caption; please state that ↑ and ↓ denote the direction that is preferable.

Circularity Check

0 steps flagged

No significant circularity: MMOOC's evaluation is grounded in human-verified labels and disjoint judge models, with no derived quantity reducing to its own inputs.

full rationale

MMOOC is a benchmark construction and evaluation study rather than a derivation from first principles. The central empirical claim—that current MLLMs struggle to balance robust answering and appropriate refusal—is measured by Accuracy, Refusal Rate, and rationality scores (Eqs. 1–6) against labels that were filtered by unanimous MLLM agreement and then manually reviewed: “All retained samples are then manually reviewed. During human verification, annotators examine each image–question pair and verify: (1) whether the answerability judgment and reference answer are correct...” The rationality judges are explicitly disjoint from data creation and from the evaluated models: “we use GPT-5.6, Claude Opus 5, and DeepSeek-V4-Pro as independent judges, none of which participates in data generation, filtering, or the evaluated model set,” with reported human–judge agreement of 96.78% for refusal rate. No parameter is fitted to a subset of the data and then reported as a prediction; the OOC categories are adopted from external work (MoHoBench, UPD) rather than from a self-citation chain, and the self-citations in Related Work ([30], [31], [33]–[35]) are topical background, not load-bearing premises. The Appendix’s question-only analysis explicitly cautions that high refusal without the image may reflect linguistic cues, so the paper does not overclaim. The residual concern—that GPT-4o/o1/o3 took part in data filtering and also appear in the evaluated model set—is a benchmark-contamination/validity issue, not circularity of derivation: final labels were human-verified, and the judged model responses were not used to define those labels. No quoted equation or construction step reduces to its own input.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or formal entities. Its core assumptions are about label quality, judge validity, and taxonomy exhaustiveness, all of which are partially supported (human verification, human-judge agreement) but not fully quantified. Design choices like unanimity filtering and judge aggregation act as hand-set parameters.

free parameters (4)
  • Unanimity threshold in answerability filtering = 3/3 model agreement
    Samples retained only when all three judge models (GPT-4o, o1, o3) agree on answerability; threshold chosen by hand with no sensitivity analysis (Sec. 3.3).
  • Judge panel size and aggregation = 3 models, averaged
    Rationality scores use three LLM judges averaged; number and weighting are design choices (Sec. 3.5).
  • Rationality score granularity = {0, 0.25, 0.5, 0.75, 1.0}
    Discrete scoring scale chosen by hand; no evidence that this scale preserves ordinal distinctions (Sec. 3.5).
  • SFT/DPO training hyperparameters = not reported
    Post-training experiments described without training data, epochs, learning rates, or data mixtures (Table 5).
axioms (4)
  • domain assumption The unanimous answerability judgment of GPT-4o, o1, and o3 is a reliable proxy for true answerability.
    Section 3.3: samples retained only when all three models agree; no external validation of label correctness beyond human review.
  • domain assumption Human annotators' answerability and reference-answer labels are correct.
    Section 3.3 human verification; no inter-annotator agreement scores reported for the final dataset.
  • domain assumption LLM-as-a-judge rationality scores are a valid measure of reasoning quality.
    Section 3.5; validated on a 2K-response subset (89.34% agreement with humans) but applied to all responses without per-category calibration.
  • domain assumption MMOOC's OOC and shifted-IC categories are mutually exclusive and jointly exhaustive.
    Section 3.1 assigns each sample to a single category; no analysis of boundary cases, category overlap, or label noise.

pith-pipeline@v1.3.0-alltime-deepseek · 26292 in / 11571 out tokens · 122366 ms · 2026-08-04T01:18:45.218883+00:00 · methodology

0 comments
read the original abstract

Multimodal Large Language Models (MLLMs) have achieved strong performance on a wide range of vision-language tasks, but often fail under imperfect or shifted contexts. A reliable MLLM should refuse truly out-of-context (OOC) questions with subject-level context shifts while still answering shifted in-context (Shifted IC) questions with non-subject context shifts. Existing benchmarks mainly target OOC or visually unanswerable questions, but overlook answerable Shifted IC cases and cover limited OOC shifts. To fill this gap, we present MMOOC, a large-scale benchmark for evaluating refusal and robust answering abilities of MLLMs. MMOOC contains over 41K image-question pairs, including answerable Shifted IC cases and unanswerable OOC cases, spanning three question formats, eight shift types and six visual scenarios, with data quality ensured through MLLM-based filtering and human verification. We evaluate model responses using Accuracy and Refusal Rate, and further introduce an LLM-as-a-Judge metric to assess the correctness of model reasoning. Experiments on diverse MLLMs show that current models still struggle to balance answer-ability and refusal under shifted contexts. We further analyze key failure patterns and show that post-training can improve robustness. MMOOC will be made publicly available.

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

80 extracted references · 21 linked inside Pith

  1. [1]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    JoshAchiam,StevenAdler,SandhiniAgarwal,LamaAhmad,IlgeAkkaya,FlorenciaLeoniAleman,DiogoAlmeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    2023

  3. [3]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024

    2024

  6. [6]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Sniffer: Multimodal large language model for explainable out-of-context misinformation detection

    Peng Qi, Zehong Yan, Wynne Hsu, and Mong Li Lee. Sniffer: Multimodal large language model for explainable out-of-context misinformation detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13052–13062, 2024

    2024

  8. [8]

    Unsolvable problem detection: Robust understanding evaluation for large multimodal models

    Atsuyuki Miyai, Jingkang Yang, Jingyang Zhang, Yifei Ming, Qing Yu, Go Irie, Yixuan Li, Hai Helen Li, Ziwei Liu, and Kiyoharu Aizawa. Unsolvable problem detection: Robust understanding evaluation for large multimodal models. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6497–6540, 2025

    2025

  9. [9]

    Haloquest: Avisual hallucination dataset for advancing multimodal reasoning

    ZhecanWang, GarrettBingham, AdamsWeiYu, QuocVLe, ThangLuong, andGolnazGhiasi. Haloquest: Avisual hallucination dataset for advancing multimodal reasoning. InEuropean Conference on Computer Vision, pages 288–304. Springer, 2024

    2024

  10. [10]

    Mohobench: Assessing honesty of multimodal large language models via unanswerable visual questions

    YanxuZhu,ShitongDuan,XiangxuZhang,JitaoSang,PengZhang,TunLu,XiaoZhou,JingYao,XiaoyuanYi,and Xing Xie. Mohobench: Assessing honesty of multimodal large language models via unanswerable visual questions. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 29205–29213, 2026

    2026

  11. [11]

    Clip-up: Clip-based unanswerable problem detection for visual question answering

    Ben Vardi, Oron Nir, and Ariel Shamir. Clip-up: Clip-based unanswerable problem detection for visual question answering. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5898–5908, 2026

    2026

  12. [12]

    Evaluating object hallucination in large vision-language models

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. InProceedings of the 2023 conference on empirical methods in natural language processing, pages 292–305, 2023. Visual Computing Lab·The Hong Kong Polytechnic University 10 / 25

    2023

  13. [13]

    Hallucination of multimodal large language models: A survey.arXiv preprint arXiv:2404.18930, 2024

    ZechenBai,PichaoWang,TianjunXiao,TongHe,ZongboHan,ZhengZhang,andMikeZhengShou. Hallucination of multimodal large language models: A survey.arXiv preprint arXiv:2404.18930, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Mirage: Assessing hallucination in multimodal reasoning chains of mllm.Advances in Neural Information Processing Systems, 38: 122910–122955, 2026

    Bowen Dong, Minheng Ni, Zitong Huang, Guanglei Yang, Wangmeng Zuo, and Lei Zhang. Mirage: Assessing hallucination in multimodal reasoning chains of mllm.Advances in Neural Information Processing Systems, 38: 122910–122955, 2026

    2026

  15. [15]

    the moon is made of marshmallows

    Yifei Ming, Senthil Purushwalkam, Shrey Pandit, Zixuan Ke, Xuan-Phi Nguyen, Caiming Xiong, and Shafiq Joty. Faitheval: Can your language model stay faithful to context, even if" the moon is made of marshmallows".arXiv preprint arXiv:2410.03727, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Felm: Benchmarking factuality evaluation of large language models.Advances in Neural Information Processing Systems, 36:44502–44523, 2023

    Yiran Zhao, Jinghan Zhang, I Chern, Siyang Gao, Pengfei Liu, Junxian He, et al. Felm: Benchmarking factuality evaluation of large language models.Advances in Neural Information Processing Systems, 36:44502–44523, 2023

    2023

  17. [17]

    Survey on factuality in large language models.ACM Computing Surveys, 58 (1):1–37, 2025

    Cunxiang Wang, Xiaoze Liu, Yuanhao Yue, Qipeng Guo, Xiangkun Hu, Xiangru Tang, Tianhang Zhang, Cheng Jiayang, Yunzhi Yao, Xuming Hu, et al. Survey on factuality in large language models.ACM Computing Surveys, 58 (1):1–37, 2025

    2025

  18. [18]

    Drift: Enhancing llm faithfulness in rationale generation via dual-reward probabilistic inference

    Jiazheng Li, Hanqi Yan, and Yulan He. Drift: Enhancing llm faithfulness in rationale generation via dual-reward probabilistic inference. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6850–6866, 2025

    2025

  19. [19]

    Is factuality enhancement a free lunch for llms? better factuality can lead to worse context-faithfulness.arXiv preprint arXiv:2404.00216, 2024

    Baolong Bi, Shenghua Liu, Yiwei Wang, Lingrui Mei, Junfeng Fang, Hongcheng Gao, Shiyu Ni, and Xueqi Cheng. Is factuality enhancement a free lunch for llms? better factuality can lead to worse context-faithfulness.arXiv preprint arXiv:2404.00216, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Mmooc: A multimodal misinformation dataset for out-of-context news analysis

    Qingzheng Xu, Heming Du, Huiqiang Chen, Bo Liu, and Xin Yu. Mmooc: A multimodal misinformation dataset for out-of-context news analysis. InAustralasian Conference on Information Security and Privacy, pages 444–459. Springer, 2024

    2024

  21. [21]

    Mfc-bench: Benchmarking multimodal fact-checking with large vision-language models.arXiv preprint arXiv:2406.11288, 2024

    Shengkang Wang, Hongzhan Lin, Ziyang Luo, Zhen Ye, Guang Chen, and Jing Ma. Mfc-bench: Benchmarking multimodal fact-checking with large vision-language models.arXiv preprint arXiv:2406.11288, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Reversible primitive– composition alignment for continual vision–language learning

    Canran Xiao, Tianxiang Xu, Siyuan Ma, Yiyang Jiang, Haoyu Gao, and Yuhan Wu. Reversible primitive– composition alignment for continual vision–language learning. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  23. [23]

    Comem: Compositional concept-graph memory for vision–language adaptation

    Heng Zhou, Jing Tang, Jusheng Zhang, Yanshu Li, Canran Xiao, Liwei Hou, Zong Ke, and Jiawei Yao. Comem: Compositional concept-graph memory for vision–language adaptation. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  24. [24]

    Attention-spectrum regularization for replay-free continual multimodal llms.arXiv preprint arXiv:2606.23063, 2026

    Chuangxin Zhao, Canran Xiao, Siyuan Ma, Mengyao Lyu, Yanbiao Ma, Jun Xia, Guiguang Ding, and Yang Liu. Attention-spectrum regularization for replay-free continual multimodal llms.arXiv preprint arXiv:2606.23063, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    Dismantling pathological shortcuts: A causal framework for faithful lvlm decoding.arXiv preprint arXiv:2606.27596, 2026

    Liu Yu, Can Chen, Ping Kuang, Zhikun Feng, Fan Zhou, and Gillian Dobbie. Dismantling pathological shortcuts: A causal framework for faithful lvlm decoding.arXiv preprint arXiv:2606.27596, 2026

    Pith/arXiv arXiv 2026

  26. [26]

    Causally-grounded dual-path attention intervention for object hallucination mitigation in lvlms

    Liu Yu, Zhonghao Chen, Ping Kuang, Zhikun Feng, Fan Zhou, Lan Wang, and Gillian Dobbie. Causally-grounded dual-path attention intervention for object hallucination mitigation in lvlms. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 36021–36029, 2026

    2026

  27. [27]

    A survey on the honesty of large language models.arXiv preprint arXiv:2409.18786, 2024

    Siheng Li, Cheng Yang, Taiqiang Wu, Chufan Shi, Yuji Zhang, Xinyu Zhu, Zesen Cheng, Deng Cai, Mo Yu, Lemao Liu, et al. A survey on the honesty of large language models.arXiv preprint arXiv:2409.18786, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Alignment for honesty.Advances in Neural Information Processing Systems, 37:63565–63598, 2024

    Yuqing Yang, Ethan Chern, Xipeng Qiu, Graham Neubig, and Pengfei Liu. Alignment for honesty.Advances in Neural Information Processing Systems, 37:63565–63598, 2024

    2024

  29. [29]

    Honestllm: Toward an honest and helpful large language model.Advances in Neural Information Processing Systems, 37:7213–7255, 2024

    Gao Chujie, Siyuan Wu, Yue Huang, Dongping Chen, Qihui Zhang, Zhengyan Fu, Yao Wan, Lichao Sun, and Xiangliang Zhang. Honestllm: Toward an honest and helpful large language model.Advances in Neural Information Processing Systems, 37:7213–7255, 2024

    2024

  30. [30]

    Dual memory networks: A versatile adaptation approach for vision-language models

    Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 28718–28728, 2024. Visual Computing Lab·The Hong Kong Polytechnic University 11 / 25

    2024

  31. [31]

    Lapt: Label-driven automated prompt tuning for ood detection with vision-language models

    Yabin Zhang, Wenjie Zhu, Chenhang He, and Lei Zhang. Lapt: Label-driven automated prompt tuning for ood detection with vision-language models. InEuropean conference on computer vision, pages 271–288. Springer, 2024

    2024

  32. [32]

    Sharpness-aware gradient matching for domain generalization

    Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3769–3778, 2023

    2023

  33. [33]

    Ants: Adaptive negative textual space shaping for ood detection via test-time mllm understanding and reasoning.arXiv preprint arXiv:2509.03951, 2025

    Wenjie Zhu, Yabin Zhang, Xin Jin, Wenjun Zeng, and Lei Zhang. Ants: Adaptive negative textual space shaping for ood detection via test-time mllm understanding and reasoning.arXiv preprint arXiv:2509.03951, 2025

    arXiv 2025

  34. [34]

    Knowledge regularized negative feature tuning of vision-language models for out-of-distribution detection

    Wenjie Zhu, Yabin Zhang, Xin Jin, Wenjun Zeng, and Lei Zhang. Knowledge regularized negative feature tuning of vision-language models for out-of-distribution detection. InProceedings of the 33rd ACM International Conference on Multimedia, pages 3565–3574, 2025

    2025

  35. [35]

    Dual distribution estimation for zero-shot noisy test-time adaptation with vlms.arXiv preprint arXiv:2606.25758, 2026

    Wenjie Zhu, Yabin Zhang, Liang Xu, Xin Jin, Wenjun Zeng, and Lei Zhang. Dual distribution estimation for zero-shot noisy test-time adaptation with vlms.arXiv preprint arXiv:2606.25758, 2026

    Pith/arXiv arXiv 2026

  36. [36]

    Combiner: Composed image retrieval guided by attribute-based neighbor relations.IEEE Transactions on Image Processing, 2026

    Zixu Li, Yupeng Hu, Zhiwei Chen, Haokun Wen, Xuemeng Song, and Liqiang Nie. Combiner: Composed image retrieval guided by attribute-based neighbor relations.IEEE Transactions on Image Processing, 2026

    2026

  37. [37]

    Conesep: Cone-based robust noise-unlearning compositional network for composed image retrieval.arXiv preprint arXiv:2604.20358, 2026

    Zixu Li, Yupeng Hu, Zhiwei Chen, Mingyu Zhang, Zhiheng Fu, and Liqiang Nie. Conesep: Cone-based robust noise-unlearning compositional network for composed image retrieval.arXiv preprint arXiv:2604.20358, 2026

    Pith/arXiv arXiv 2026

  38. [38]

    Hybridmeta-learnersforestimatingheterogeneoustreatment effects.arXiv preprint arXiv:2506.13680, 2025

    ZhongyuanLiang,LarsvanderLaan,andAhmedAlaa. Hybridmeta-learnersforestimatingheterogeneoustreatment effects.arXiv preprint arXiv:2506.13680, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Oc-distill: Ontology-aware contrastive learning with cross-modal distillation for icu risk prediction.arXiv preprint arXiv:2604.16878, 2026

    Zhongyuan Liang, Junhyung Jo, Hyang-Jung Lee, Sang Kyu Kim, and Irene Y Chen. Oc-distill: Ontology-aware contrastive learning with cross-modal distillation for icu risk prediction.arXiv preprint arXiv:2604.16878, 2026

    Pith/arXiv arXiv 2026

  40. [40]

    Or-bench: Anover-refusalbenchmarkforlargelanguage models.arXiv preprint arXiv:2405.20947, 2024

    JustinCui, Wei-LinChiang, IonStoica, andCho-JuiHsieh. Or-bench: Anover-refusalbenchmarkforlargelanguage models.arXiv preprint arXiv:2405.20947, 2024

    Pith/arXiv arXiv 2024

  41. [41]

    Abstentionbench: Reasoning llms fail on unanswerable questions.arXiv preprint arXiv:2506.09038, 2025

    Polina Kirichenko, Mark Ibrahim, Kamalika Chaudhuri, and Samuel J Bell. Abstentionbench: Reasoning llms fail on unanswerable questions.arXiv preprint arXiv:2506.09038, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Refusalbench: Generative evaluation of selective refusal in grounded language models

    AashiqMuhamed,LeonardoFRRibeiro,MarkusDreyer,VirginiaSmith,andMonaDiab. Refusalbench: Generative evaluation of selective refusal in grounded language models. InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6811–6856, 2026

    2026

  43. [43]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744, 2022

    2022

  44. [44]

    Discrepancy learning guided hierarchical fusion network for multi-modal recommendation.Knowledge-Based Systems, 317:113496, 2025

    Yuzhuo Dang, Zhiqiang Pan, Xin Zhang, Wanyu Chen, Fei Cai, and Honghui Chen. Discrepancy learning guided hierarchical fusion network for multi-modal recommendation.Knowledge-Based Systems, 317:113496, 2025

    2025

  45. [45]

    Mappo: Maximum a posteriori preference optimization with prior knowledge.arXiv preprint arXiv:2507.21183, 2025

    Guangchen Lan, Sipeng Zhang, Tianle Wang, Yuwei Zhang, Daoan Zhang, Xinpeng Wei, Xiaoman Pan, Hongming Zhang, Dong-Jun Han, and Christopher G Brinton. Mappo: Maximum a posteriori preference optimization with prior knowledge.arXiv preprint arXiv:2507.21183, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    Direct preferenceoptimization: Yourlanguagemodelissecretlyarewardmodel.Advancesinneuralinformationprocessing systems, 36:53728–53741, 2023

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preferenceoptimization: Yourlanguagemodelissecretlyarewardmodel.Advancesinneuralinformationprocessing systems, 36:53728–53741, 2023

    2023

  47. [47]

    Simpo: Simple preference optimization with a reference-free reward

    Yu Meng, Mengzhou Xia, and Danqi Chen. Simpo: Simple preference optimization with a reference-free reward. Advances in Neural Information Processing Systems, 37:124198–124235, 2024

    2024

  48. [48]

    Sppo: Sequence-level ppo for long-horizon reasoning tasks, 2026

    Tianyi Wang, Yixia Li, Long Li, Yibiao Chen, Shaohan Huang, Yun Chen, Peng Li, Yang Liu, and Guanhua Chen. Sppo: Sequence-level ppo for long-horizon reasoning tasks, 2026. URL https://arxiv.org/abs/2604.08865

    Pith/arXiv arXiv 2026

  49. [49]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  50. [50]

    Negative object presence evaluation (nope) to measure object hallucination in vision-language models

    Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. InProceedings of the 3rd Workshop on Advances in Language and Vision Research (ALVR), pages 37–58, 2024. Visual Computing Lab·The Hong Kong Polytechnic University 12 / 25

    2024

  51. [51]

    Mme: A comprehensive evaluation benchmark for multimodal large language models.arXiv preprint arXiv:2306.13394, 2023

    Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models.arXiv preprint arXiv:2306.13394, 2023

    Pith/arXiv arXiv 2023

  52. [52]

    Are we on the right way for evaluating large vision-language models?Advances in Neural Information Processing Systems, 37:27056–27087, 2024

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models?Advances in Neural Information Processing Systems, 37:27056–27087, 2024

    2024

  53. [53]

    Ok-vqa: A visual question answering benchmarkrequiringexternalknowledge

    Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmarkrequiringexternalknowledge. InProceedingsoftheIEEE/cvfconferenceoncomputervisionandpattern recognition, pages 3195–3204, 2019

    2019

  54. [54]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean conference on computer vision, pages 740–755. Springer, 2014

    2014

  55. [55]

    Vqa: Visual question answering

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. InProceedings of the IEEE international conference on computer vision, pages 2425–2433, 2015

    2015

  56. [56]

    Gqa: A new dataset for real-world visual reasoning and compositional question answering

    Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700–6709, 2019

    2019

  57. [57]

    Mmbench: Is your multi-modal model an all-around player? InEuropean conference on computer vision, pages 216–233

    YuanLiu,HaodongDuan,YuanhanZhang,BoLi,SongyangZhang,WangboZhao,YikeYuan,JiaqiWang,Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? InEuropean conference on computer vision, pages 216–233. Springer, 2024

    2024

  58. [58]

    Isthis<subject/description>?

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9556–9567, 2024. Visual Computing L...

    2024

  59. [60]

    Is this...?

    Write TWO questions in “Is this...?” Yes/No format, each embedding a FALSE, UNCERTAIN, or IRRELEVANTpremiseinthequestionstem(e.g.,“IsthisthecapitalcityofFrance,giventhatitwasbuilt in 1800?” — when the image shows something unrelated). The premise must be incorrect or unverified, yet the core Yes/No ask must still be answerable from the image. • Question 1...

  60. [61]

    yes” or “no

    For each question, provide the correct answer (“yes” or “no”) based on what is actually shown

  61. [62]

    Is this...?

    For each question, provide reasoning explaining: (a) why the premise is false/unsupported, and (b) why the model can still confidently answer the core Yes/No question. Output format: Return a JSON array with exactly 2 objects. Each object must have these keys: •caption: brief description of what’s in the image •question: the yes/no question in “Is this......

  62. [64]

    Based on X which is not shown

    Write ONE multi-choice question with exactly 4 options (A, B, C, D) that CAN be confidently answered by looking at the image, but whose question stem includes a FALSE, UNCERTAIN, or IRRELEVANT premise (e.g., “Based on X which is not shown...”, “Since Y happened...”, “Given that Z is true...”). The premise must be incorrect or unverified, yet the core ask ...

  63. [65]

    Provide the correct option letter (A, B, C, or D)

  64. [66]

    A”, “B”, “C

    Provide reasoning explaining: (a) why the premise is false/unsupported, and (b) why the model can still answer the core question. Output format: Return a JSON array with exactly 1 object. Each object must have these keys: •caption: brief description of what’s in the image •question: the multi-choice question with ALL four options (A. B. C. D.) included in...

  65. [68]

    Basedonthispainting by Picasso from 1920, what object is in the center?

    Write ONE open-ended question that CAN be confidently answered by looking at the image, but whose questionstemincludesaFALSE,UNCERTAIN,orIRRELEVANTpremise(e.g., “Basedonthispainting by Picasso from 1920, what object is in the center?” — when the image shows a photograph, not a Picasso painting). The premise must be incorrect or unverified, yet the core as...

    1920

  66. [70]

    misleading_premise_ic

    Provide reasoning explaining: (a) why the premise is false/unsupported, and (b) why the model can still answer the core question. Output format: Return a JSON array with exactly 1 object. Each object must have these keys: •caption: brief description of what’s in the image •question: the open-ended VQA question with a false premise embedded in the stem •an...

  67. [72]

    What is the person doing and what is the weather like?

    Write ONE open-ended question that CAN be confidently answered by looking at the image, but which contains ADDITIONAL SUB-QUESTIONS or PARTS that CANNOT be answered from the image (e.g., “What is the person doing and what is the weather like?” — where the person’s action is visible but the weather cannot be determined). The question must still have a clea...

  68. [74]

    partially_answerable_ic

    Provide reasoning explaining: (a) which part of the question is unanswerable from the image, and (b) why the answerable part is still confidently derivable. Output format: Return a JSON array with exactly 1 object. Each object must have these keys: •caption: brief description of what’s in the image •question : the open-ended VQA question containing both a...

  69. [75]

    Write a brief caption describing what is ACTUALLY visible in the image

  70. [76]

    At this formal dinner event, what is on the table?

    Write ONE open-ended question that CAN be confidently answered by looking at the image, but which is framed with a MISMATCHING scene description or context (e.g., “At this formal dinner event, what is on the table?” — when the image actually shows a beach scene). The image and text context disagree at the scene/narrative level, yet specific visual facts i...

  71. [77]

    Provide the correct answer that can be directly derived from the image

  72. [78]

    image_text_mismatch_ic

    Provide reasoning explaining: (a) how the provided context mismatches the image, and (b) why the target visual facts are still confidently answerable from the image. Output format: Return a JSON array with exactly 1 object. Each object must have these keys: •caption: brief description of what is ACTUALLY in the image (without the mismatching context) •que...

  73. [82]

    What is written on the sign?

    Multimodal Ambiguity: The problem lies in the inherent lack of clarity within the image pixels or the ambiguity in the textual query. The model cannot answer because the visual information is insufficient or the question is too vague to target a specific entity. Sub-types include (randomly select ONE sub-type for each question): • Sub-type 1: Insufficient...

  74. [86]

    What is the dog holding?

    Visual False Premises: The question is based on a false premise that contradicts what is actually shown intheimage. Thequestionassumessomethingexists(anobject,aproperty,orastate)thatisNOTpresent in the image. The model cannot answer because the premise itself is invalid—the entity or condition being asked about does not exist in the image. Sub-types inclu...

  75. [90]

    Which cup is behind the plate?

    Uncertain Spatial & Physical Context: The image does not provide sufficiently reliable cues to determine relative position, depth, orientation, occlusion, or 3D structure, or to infer stability, support, motion, or likely outcomes. Sub-types include (randomly select ONE sub-type for each question): • Sub-type 1: Uncertain Spatial Relations - The image doe...

  76. [94]

    I can’t answer

    Unclear Logical & Symbolic: The problem lies in symbols, formulas, labels, arrows, markers, or diagram elements that cannot be reliably recognized or interpreted, or in missing rules, conditions, or intermediate clues needed for logical inference. Sub-types include (randomly select ONE sub-type for each question): • Sub-type 1: Unclear Visual Symbols - Sy...

  77. [95]

    Write a brief caption describing what is visible in the image

  78. [96]

    Write ONE open-ended question that the model cannot definitively answer

  79. [97]

    Categories and question types:

    For each question, provide reasoning explaining WHY the model cannot answer it. Categories and question types:

  80. [98]

    I can’t answer

    MissingKnowledge&Background: Theproblemliesininformationthatisnotintheimageandcannot be derived from it. The question requires specialized, rare, or long-tail knowledge, or cultural, historical, situational, temporal, or social background not present in or derivable from the image alone. Sub-types include (randomly select ONE sub-type for each question): ...