Pith. sign in

REVIEW 4 major objections 6 minor 18 references

Seeing vs. Believing: Evaluating the Language Bias of Open-Source MLLMs in Counter-Intuitive Scenes

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that open-source multimodal models score at chance on images of counter-intuitive actions because they override visible evidence with statistically common language, and that targeted fine-tuning can partially correct this.

desk verdict The benchmark idea is a useful extension of compositionality testing, but the paper's headline claims are unsupported by its own experiments and the reporting is too inconsistent to cite. read the letter →

arxiv 2601.07737 v2 pith:UGZFGCY6 submitted 2026-01-12 cs.CV cs.AI

Chen Ling , Tongwei Zhang , Hanqian Li , Nai Ding This is my paper · ORCID
classification cs.CVcs.AI
keywords multimodallargelanguagemodelspriorvisualgroundingcounter-intuitivescenessemanticrolereversalquestionansweringbenchmarkchain-of-thoughtreasoning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that open-source instruction-tuned multimodal models do not genuinely ground action roles in images: when shown synthetic scenes that contradict everyday common sense, such as a rabbit dragging a tiger, they choose the statistically common text description at chance level instead of describing what is actually visible. Humans score near-perfectly on the same choices, and leading proprietary models score well below humans but far above chance. The authors argue the dominant failure is a language prior—models override anomalous visual input with frequent text patterns—and they demonstrate two partial remedies: chain-of-thought prompting, which improves accuracy but slows responses and introduces a refusal-to-believe failure mode, and targeted fine-tuning, which lifts a small open-source model from chance to modest accuracy. A sympathetic reader would care because the result isolates where current open-source multimodal systems stop seeing and start guessing.

What carries the argument

The load-bearing object is the benchmark itself (CAIT/UAIT): 400 synthetic scenes generated by text-to-image diffusion from hand-verified 'uncommon-sense' sentences. Sentences are built by selecting verbs whose meaning reverses when agent and patient swap, generating common sentences with a language model, swapping the roles, and re-generating detailed visual descriptions, then filtering the generated images manually for clarity. Each image is paired with a two-option question in which the only difference between options is the role assignment, so the task isolates semantic-role grounding from object recognition and vocabulary knowledge.

What would settle it

Show the 400 raw generated images to humans in a forced-choice between 'a rabbit drags a tiger' and 'a tiger drags a rabbit' without revealing which is common; if human agreement with the intended label is substantially below the reported ~0.96, the premise that the images unambiguously depict the counter-intuitive action collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that the near-chance performance of open-source MLLMs on counter-intuitive scenes is a language-prior failure, not a vision failure. It constructs a benchmark of 400 synthetic images in which the correct caption differs from the common caption only by swapping agent and patient roles (e.g., 'a rabbit is dragging a tiger' versus 'a tiger is dragging a rabbit'), then asks models to pick the matching caption. Accuracy at chance on this forced choice means the model is not using the visible role relation. On the paper's own results, humans reach about 0.96, proprietary models reach up to about 0.88, and standard open-source instruction-tuned models score at or near c

Load-bearing premise

The benchmark's validity rests on the assumption that the synthetic images unambiguously depict the swapped roles, so that chance-level model performance is a reasoning failure rather than a failure of the image generator to produce the intended scene.

Editorial extensions

If this is right

  • If the central claim holds, standard open-source MLLMs cannot be reliably used for visual reasoning in rare, anomalous, or safety-relevant scenes where the common-sense expectation and the actual visual evidence diverge.
  • The benchmark provides a minimal diagnostic: chance-level forced choice on role-swapped captions is sufficient to detect a language prior, so model releases could be screened this way.
  • Chain-of-thought reasoning is not a cure: the paper shows it can induce overthinking that rejects valid visual evidence, meaning reasoning prompts can trade one failure for another.
  • Fine-tuning on counter-intuitive data substantially reduces the language prior in a small model, suggesting data-level mitigation is feasible at modest compute.
  • Contrastive image-text models also fail role reversals, situating the failure as a general multimodal alignment problem rather than an instruction-tuning artifact.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The role-swap design could be reused as a probe in other multimodal systems: feeding the same image with swapped captions through different vision encoders would test whether the language prior originates in the language decoder or in the vision-language alignment stage.
  • The overthinking failure mode suggests a reverse-hallucination tendency: models can be trained to trust internal world models over pixels, which is a distinct failure from the usual case of trusting language priors; safety-critical deployments should test both directions.
  • Extending the benchmark to multi-lingual or culturally-specific counter-intuitive scenes could reveal whether the language prior is an English-corpus artifact or a general property of alignment training.
  • The reported fine-tuning gain on the benchmark but smaller gain on a related compositionality test hints that the mitigation may be task-specific; a useful follow-up is measuring whether the improvement transfers to other role-reversal diagnostics.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces UAIT, a benchmark of 400 synthetic counter-intuitive action scenes (e.g., 'a rabbit is dragging a tiger') and evaluates vision-language models (VLMs) and contrastive-learning models on a two-choice VQA task. The abstract claims that humans reach ~0.95 accuracy, leading proprietary models achieve up to 0.88, 14 representative open-source instruction-tuned models perform at chance level, and that failures are driven by a language prior that overrides visual evidence; it further claims that Chain-of-Thought prompting slows responses and creates an 'overthink/refuse' failure mode, while fine-tuning mitigates the bias. The body, however, reports results for only three open-source VLMs (Qwen2-VL-Instruct 0.64, LLaVA1.5 0.36, LLaMA3.2-Vision 0.69), two contrastive models, and two human annotators; no proprietary models are evaluated, and no control isolating a language prior is presented.

Significance. If the benchmark were validated and the empirical claims supported, UAIT could be a useful diagnostic tool for evaluating semantic-role understanding and common-sense reasoning in VLMs. The semi-automated construction pipeline (VerbNet-based verb selection, LLM-assisted text generation, Stable Diffusion synthesis with manual screening) is a constructive contribution, and the LoRA fine-tuning demonstration is a relevant exploratory result. However, the current manuscript's evidence does not support the headline conclusions: the model coverage is far narrower than claimed, the reported numbers are internally inconsistent, and the lack of image-validity and language-prior controls means the benchmark's validity is not yet established. These are load-bearing issues rather than presentational ones.

major comments (4)
  1. [Abstract vs. §4.1, Table 5] The abstract's central claim — that 14 open-source MLLMs perform at chance and that proprietary models (Claude, Gemini) reach up to 0.88 — is unsupported by the body. Only Qwen2-VL-Instruct, LLaVA1.5, and LLaMA3.2-Vision are evaluated, and no proprietary model appears anywhere. Moreover, the reported 0.64 and 0.69 for Qwen2 and LLaMA3.2 are not chance-level: with N=400, the binomial standard error is about 0.025, placing these values roughly 5–8 standard errors above 0.5. Only LLaVA1.5 (0.36) is below chance. The abstract's blanket 'chance level' characterization is therefore contradicted by the paper's own data.
  2. [Table 5 vs. Table 7, §4.2 vs. Table 6] There are internal numeric inconsistencies that undermine the reliability of the reported results. The accuracies for CLIP and RWKV-CLIP are 0.49/0.53 in Table 5 but are swapped to 0.53/0.49 in Table 7. In addition, LLaVA1.5's baseline accuracy is 0.36 in Table 5 but 0.40 in Table 6, and the fine-tuned LLaVA-UAIT model is reported as 0.79 in the §4.2 text but 0.85 in Table 6. These discrepancies mean a reader cannot determine which numbers are authoritative.
  3. [§3.3, §3.4, §4.2] The paper attributes model errors to a 'strong language prior' overriding visual evidence, but no experiment isolates a language prior. There is no text-only baseline (answering from the options with no image), no image-only or human forced-choice validation on the raw generated images, and no inter-annotator agreement metric for the manual image screening. As the authors acknowledge in §3.3, screening was manual; with only two annotators for the human benchmark and no agreement measure, the possibility that low model accuracy reflects Stable Diffusion generation artifacts or ambiguous images rather than language bias is not excluded. This is a load-bearing gap for the paper's main interpretation.
  4. [§3.4, §4.2, Abstract] The Chain-of-Thought claims are internally inconsistent and unsupported. §3.4 states that the prompt ends with 'Output only the corresponding letter without providing a reason' to eliminate CoT effects, yet §4.2 describes the Qwen2 and LLaMA3.2 results as being obtained 'after enabling Chain-of-Thought (CoT) reasoning.' The abstract further claims that CoT 'significantly slows down the response and generates a new failure mode: models overthink and refuse to accept the actual visual content.' No experiment reports response-time measurements, refusal rates, or a systematic CoT-vs-no-CoT comparison. These claims appear in the abstract but are not evidenced in the results.
minor comments (6)
  1. [Throughout] Numerous typographical errors: 'Relation Work' should be 'Related Work'; 'Vison and Langugae models' should be 'Vision and Language models'; 'difficulty' and 'efficiency' contain ligature issues; 'saluate' should be 'salute'. A careful proofread is needed.
  2. [References [17]] Reference [17] is cited for LLaMA3.2-Vision but refers to LLaMA-Adapter V2 (Gao et al., 2023). This is not the correct source for the LLaMA3.2-Vision model; please cite the appropriate model release or technical report.
  3. [Title] The arXiv title 'Seeing vs. Believing: Evaluating the Language Bias of Open-Source MLLMs in Counter-Intuitive Scenes' differs substantially from the document's internal title 'Evaluating the encoding competence of visual language models using uncommon actions.' Ensure a single consistent title that reflects the actual content.
  4. [Table 1] The column 'Judgement Basis' is unclear (e.g., 'direction of agent', 'active, passive, body contact') and the table does not define the meaning of the numeric classes. Provide a legend or a short explanation so readers can interpret the verb library composition.
  5. [Table 5, Table 7] In addition to the swapped values, neither table includes confidence intervals or the number of items per model. Given N=400 and the inconsistent numbers, confidence intervals are essential for comparing models and for assessing the 'chance-level' claim.
  6. [Figure 5] The LoRA fine-tuning loss curve is never referenced in the text and its placement is unclear. Either reference it in §4.2/§4.5 or remove it.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the benchmark is empirical and its central claim, though under-supported, is not entailed by construction.

full rationale

I walked the paper's derivation chain looking for the listed circularity patterns. The benchmark is constructed by taking common sentences, swapping agent and patient, generating images from the swapped descriptions, and asking models to choose between the common and uncommon captions. The 'uncommon' label is defined by the swap (Sec. 3.2), and the correct answer is the generation prompt, but the paper does not fit any parameter to the evaluation outputs and then rename it as a prediction. The 'language prior' conclusion is an interpretation of the error pattern; it is not algebraically or definitionally forced by the construction. There are no self-citations, no imported uniqueness theorem, and no ansatz smuggled in via citation. The fine-tuning experiment uses a genuine 70/30 split and is not a fitted-input-called-prediction step. What the paper lacks are controls (text-only, image-only, forced-choice human validation on raw generated images), and its abstract conflicts with Table 5 (three open-source models, not 14; proprietary models not actually reported). Those are serious empirical validity and reporting problems, but underdetermination and inconsistency are not circularity. No step reduces to its own inputs, so the honest finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's central claim rests on domain assumptions about the visual validity of generated counter-intuitive scenes and the reliability of a two-person human baseline. No new physical or conceptual entities are introduced. The main untested assumption is that the synthetic images unambiguously depict the intended role reversal.

free parameters (3)
  • Dataset size = 400
    Chosen by the authors; the held-out test set (30%) is only ~120 items, leading to wide confidence intervals for the fine-tuning claim.
  • Stable Diffusion guidance scale = 8.0
    Hand-set generation hyperparameter that affects visual faithfulness of the generated action scenes.
  • LoRA learning rate = 1e-4
    Hand-set fine-tuning hyperparameter; no sensitivity analysis is reported.
assumptions (4)
  • domain assumption The transfer rule swapping agent and patient yields a counter-common-sense scene that is visually distinguishable and semantically opposite.
    Section 3.2 states the core construction rule; if swapped scenes are not visually distinguishable (e.g., 'guide dog follows blind man' vs 'blind man follows guide dog' may look similar), the benchmark's intended signal collapses.
  • domain assumption Stable Diffusion 3.5 generated images faithfully realize the intended uncommon action with the correct agent/patient assignment.
    Section 3.3 assumes the generated 512x512 images depict the described role relationship; manual screening is used but not quantified with agreement statistics.
  • domain assumption Two untrained annotators provide a reliable human baseline (0.96).
    Section 4.1 states human performance is the average of two participants; no power analysis or inter-annotator agreement is reported.
  • ad hoc to paper The multiple-choice prompt ('Output only the corresponding letter') eliminates chain-of-thought effects.
    Section 3.4 assumes the prompt suppresses CoT, but the body reports CoT variants that change accuracy, so the assumption is questionable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Seeing vs. Believing: Evaluating the Language Bias of Open-Source MLLMs in Counter-Intuitive Scenes." pith.science (2026). https://pith.science/paper/UGZFGCY6

@misc{pith2026260107737,
  author       = {Pith},
  title        = {Pith review of: Seeing vs. Believing: Evaluating the Language Bias of Open-Source MLLMs in Counter-Intuitive Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UGZFGCY6}},
  note         = {Machine review of arXiv:2601.07737}
}
read the original abstract

Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance in mainstream visual understanding tasks, but their ability to process action scenes that contradict everyday common sense remains undertested. To address this gap, we introduce CAIT, a benchmark comprising 400 high-fidelity synthetic scenes focused on counter-intuitive visual actions, such as ``a rabbit is chasing a tiger'', where visual evidence explicitly contradicts common-sense expectations. We evaluate human, leading proprietary models (e.g., Claude and Gemini), and 14 representative open-source MLLMs. Humans achieve near-perfect performance (around 0.95 accuracy) and proprietary models demonstrate robust understanding (achieving up to 0.88 accuracy), standard open-source instruction-tuned models perform at the chance level. Further analysis demonstrates that this failure is driven by a strong language prior: rather than trusting the visual input, they automatically override the anomalous visual signals with statistically common text descriptions. Although introducing Chain-of-Thought reasoning mechanisms can improve accuracy, it significantly slows down the response and generates a new failure mode: models overthink the scenario and refuse to accept the actual visual content simply because it violates real-world physical laws. Finally, we demonstrate that targeted fine-tuning and structured prompting can effectively mitigate this reliance on language priors, enabling open-source models to accurately ground their reasoning in actual visual evidence.

Figures

Figures reproduced from arXiv: 2601.07737 by the authors.

Figure 1
Figure 1. This is an example from UAIT, the correct description of the action scene is “a rabbit is dragging a tiger”. Below are the reasons why the image is beyond the common-sense understanding. using prompt engineering and few-shot learning techniques. • Text-to-image generation for unusual visual events guided by detailed scene descrip￾tions. • Evaluation through accuracy comparison and reasoning analysis. Our work also p… view at source ↗
Figure 2
Figure 2. The construction framework of uncommon image-text To achieve this, we used the advanced text-generated image model, stable diffusion 3. Considering that directly generating ”uncommon” images may face the problems of seman￾tic ambiguity and inaccurate description, we introduced an auxiliary generation method based on detailed counter-common sense visual descriptions. In the specific operation pro￾cess, we first selec… view at source ↗
Figure 3
Figure 3. examples of a VQA tasks in the UAIT dataset. Each image corresponds to a question and two options. To eliminate the influence of CoT when the model answers, the end of the question contains ”Output only the corresponding letter of the option without providing a reason.” During the dataset construction process, we paid special attention to how to guide large 17 [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: An example of model answer on the VQA task in UAIT dataset, we use qwen2- vl-instruct as a base model here. 19 [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: lora finetuning loss of LLaVA-1.5 on UAIT 4.6.1 Stripping away contextual statistical bias Traditional models often rely on the inherent statistical laws in a large amount of cor￾pus, and habitually select character combinations that appear frequently in training. UAIT…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 13 linked inside Pith

  1. [1]

    GPT-4 Technical Report; 2024

    OpenAI, Achiam J, Adler S, Agarwal S, Ahmad L, Akkaya I, et al.. GPT-4 Technical Report; 2024. A vailable from: https://arxiv.org/abs/2303.08774

    arXiv 2024

  2. [2]

    Visual Instruction Tuning

    Liu H, Li C, Wu Q, Lee YJ. Visual Instruction Tuning. In: Oh A, Naumann T, Globerson A, Saenko K, Hardt M, Levine S, editors. Advances in Neural Information Processing Systems. vol. 36. Curran Associates, Inc.; 2023. p. 34892-916. A vail- able from: https://proceedings.neurips.cc/paper_files/paper/2023/file/ 6dcf277ea32ce3288914faf369fe6de0-Paper-Conference.pdf

    2023

  3. [3]

    Qwen2-VL: Enhancing Vision- Language Model’s Perception of the World at Any Resolution; 2024

    Wang P, Bai S, Tan S, Wang S, Fan Z, Bai J, et al.. Qwen2-VL: Enhancing Vision- Language Model’s Perception of the World at Any Resolution; 2024. A vailable from: https://arxiv.org/abs/2409.12191

    arXiv 2024

  4. [4]

    Describing Common Human Visual Actions in Images; 2015

    Ronchi MR, Perona P. Describing Common Human Visual Actions in Images; 2015. A vailable from: https://arxiv.org/abs/1506.02203

    arXiv 2015

  5. [5]

    Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality

    Thrush T, Jiang R, Bartolo M, Singh A, Williams A, Kiela D, et al. Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 2022. p. 5238-48

    2022

  6. [6]

    High-Resolution Image Synthesis with Latent Diffusion Models; 2022

    Rombach R, Blattmann A, Lorenz D, Esser P, Ommer B. High-Resolution Image Synthesis with Latent Diffusion Models; 2022. A vailable from: https://arxiv.org/ abs/2112.10752

    arXiv 2022

  7. [7]

    Attention Is All You Need; 2023

    Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al.. Attention Is All You Need; 2023. A vailable from: https://arxiv.org/abs/1706.03762

    arXiv 2023

  8. [8]

    Learning Transferable Visual Models From Natural Language Supervision; 2021

    Radford A, Kim JW, Hallacy C, Ramesh A, Goh G, Agarwal S, et al.. Learning Transferable Visual Models From Natural Language Supervision; 2021. A vailable from: https://arxiv.org/abs/2103.00020

    arXiv 2021

Show all 18 references
  1. [9]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale; 2021

    Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T, et al.. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale; 2021. A vailable from: https://arxiv.org/abs/2010.11929

    2021 arXiv

  2. [10]

    Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) network

    Sherstinsky A. Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) network. Physica D: Nonlinear Phenomena. 2020;404:132306. A vailable from: https://www.sciencedirect.com/science/ article/pii/S0167278919305974. 31 Evaluating the encoding competenc...

    2020

  3. [11]

    R WKV-CLIP: A Robust Vision- Language Representation Learner; 2024

    Gu T, Yang K, An X, Feng Z, Liu D, Cai W, et al.. R WKV-CLIP: A Robust Vision- Language Representation Learner; 2024. A vailable from: https://arxiv.org/abs/ 2406.06973

    2024 arXiv

  4. [12]

    LoRA: Low-Rank Adaptation of Large Language Models; 2021

    Hu EJ, Shen Y, Wallis P, Allen-Zhu Z, Li Y, Wang S, et al.. LoRA: Low-Rank Adaptation of Large Language Models; 2021. A vailable from: https://arxiv.org/ abs/2106.09685

    2021 arXiv

  5. [13]

    315VerbNet: Capturing English Verb Behav- ior, Meaning, and Usage

    Palmer M, Bonial C, Hwang J. 315VerbNet: Capturing English Verb Behav- ior, Meaning, and Usage. In: The Oxford Handbook of Cognitive Science. Ox- ford University Press; 2017. A vailable from: https://doi.org/10.1093/oxfordhb/ 9780199842193.013.15

    2017 doi

  6. [14]

    Qwen2 Technical Report

    Yang A, Yang B, Hui B, Zheng B, Yu B, Zhou C, et al.. Qwen2 Technical Report

  7. [15]

    Language Models are Few-Shot Learners; 2020

    Brown TB, Mann B, Ryder N, Subbiah M, Kaplan J, Dhariwal P, et al.. Language Models are Few-Shot Learners; 2020. A vailable from: https://arxiv.org/abs/ 2005.14165

    2020 arXiv

  8. [16]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models; 2023

    Wei J, Wang X, Schuurmans D, Bosma M, Ichter B, Xia F, et al.. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models; 2023. A vailable from: https://arxiv.org/abs/2201.11903

    2023 arXiv

  9. [17]

    LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model; 2023

    Gao P, Han J, Zhang R, Lin Z, Geng S, Zhou A, et al.. LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model; 2023. A vailable from: https:// arxiv.org/abs/2304.15010. 32

    2023 arXiv

  10. [2024]

    A vailable from: https://arxiv.org/abs/2407.10671

    arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.