PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluating VQA in Vision Language Models using Cooperative Principles

Monika Shah, Sudarshan Balaji, Somdeb Sarkhel, Sanorita Dey, Deepak Venugopal

arXiv:2610.02878v1cs.CL

TL;DR

The paper asks whether VLMs can handle Gricean-maxim violations in VQA, where questions include ambiguous, excessive, or false information. It uses VLM-generated question modifiers with human review and finds reduced accuracy under violations, alongside differences between human and VLM reasoning and between human-like and AI-generated violations.

  • Problem

    The study investigates how Gricean-maxim violations affect VLM VQA accuracy, whether human- and AI-generated violations elicit different reasoning, and how VLM pragmatic reasoning compares with human reasoning.

    Source: §1 · p. 1

  • Method

    The authors have VLMs add image-related modifiers to VQA questions, use humans to check modifications, and evaluate model answers to original and modified questions.

    Source: §3.1 · p. 4

  • Results

    VLM accuracy decreases with violations, models perform better on more natural human-like violations than artificial ones, and their resolution differs from human pragmatic reasoning.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The evaluated VLMs show limited pragmatic robustness to question violations, with performance depending on whether violations are human-like or AI-generated.

    Source: §1 · p. 2

    The experiments assume original human-written questions generally follow Grice’s maxims, though some may already violate them.

    Source: §7 · p. 10

Abstract

from arXiv · show

We evaluate the performance of Vision Language Models in Visual Question Answering (VQA) when questions violate Grice's maxims. To do this, we use VLMs to generate question modifiers that add non-essential, ambiguous or false information and show that in the presence of such violations, the VLMs that we evaluate (ChatGPT, Claude, Gemini and Llava) show diminished performance. Further, we empirically show the difference between how humans reason pragmatically compared to VLMs, and the difference in VLM reasoning when it resolves violations that are human-induced compared to those that are AI-generated. Finally, we show that human cognitive effort (measured through time-on-task in an experiment) is lower for resolving VLM-induced violations, but VLMs themselves perform less accurately in such cases.

1 Introduction

The paper examines whether VLMs can pragmatically resolve Gricean-maxim violations in VQA, where human listeners can infer intended meaning but face greater cognitive burden. It probes how question modifiers affect model accuracy and compares human-induced with AI-generated violations.

1 Introduction

  • The work evaluates VLM sensitivity to Cooperative-principle violations, which may make questions ambiguous, overly detailed, or contain false or irrelevant information.The authors motivate this by noting that humans can generally comprehend violations, although doing so imposes greater cognitive burden.Source: §1 · p. 1
  • VLM-generated question modifiers are human-supervised, and image context may help a model resolve a violation introduced in text.The modified questions are evaluated alongside the original questions.Source: §1 · p. 1
  • The study asks how modifiers affect VQA accuracy, whether human- and AI-generated violations produce different reasoning, and how human effort relates to VLM reasoning.The questions frame the paper’s investigation of pragmatic reasoning in VLMs.Source: §1 · pp. 1–2
  • VLM accuracy decreases with violations, models perform better on more natural human-like violations than artificial ones, and their resolution differs from human pragmatic reasoning.Experiments use ChatGPT4o, Claude Sonnet 3.5, Llava 7B, and Gemini Flash-1.5 on VQA2.0 questions.Source: §1 · p. 2

2 Related Work

Prior VQA research includes accuracy benchmarks, judge-based evaluation, hallucination datasets, and adversarial question modifications; this work extends question modification to examine VLM reasoning through cooperative principles.

2 Related Work

  • VQA research includes accuracy datasets, growing LLM-as-a-judge use, hallucination datasets, and semantically equivalent adversarial questions that can change model predictions.The cited work also includes perturbation of VQA questions to examine LXMERT; this paper broadens that modification approach to VLM reasoning grounded in cooperative principles.Source: §2 · p. 2
  • The authors build on prior VQA question-perturbation work by evaluating VLM reasoning through cooperative principles.The related work identifies Britton et al. (2022) as a study of modified VQA questions and LXMERT.Source: §2 · p. 2
  • Prior hallucination datasets include false-premise, visually challenging, insufficient-context, and synthesized image-question examples.The cited studies are Wang et al. (2024) and Wu et al. (2024).Source: §2 · p. 2

3 Evaluation Procedure

The evaluation tests VLM responses to image-related question modifiers that preserve the original answer, with humans checking modifications and comparing open-ended answers. For binary and quantitative responses, modifiers generally reduce accuracy, though the size and significance of changes vary by model and modifier source.

3 Evaluation Procedure

  • The framework treats a question as Grice-optimal when its answer can be obtained from the image, then adds image-grounded modifiers that preserve the answer.The diagnostic approach uses question modifiers to test whether response models can attend to relevant image cues and filter distractions.Source: §3 · p. 2
  • The study samples 995 VQA v2.0 test questions across binary, quantitative, and open-ended responses, using human answers as ground truth for open-ended items.The evaluated models are GPT-4o, Gemini-1.5-Flash, Claude-3.5-Sonnet, and Llava-7B.Source: §3 · p. 3
  • For binary and quantitative responses with unique ground truth, modifiers generally decrease accuracy; Claude has the smallest average reduction (4.82%), and Llava the largest (13.28%).McNemar tests found significant changes in most cases (p < 0.05); Claude’s responses were not significantly changed for three of four modification-models.Source: §3 · p. 3

3.1 Adding Modifiers

  • VLM prompts constrain modifiers to preserve the answer, question type, and context, and to add a visual or relational property without giving away the answer.Humans judge whether modifications preserve the original answer; three workers assess each, with majority judgments determining retention or correction.Source: §3 · pp. 2–3, §3.1 · p. 4
  • For open-ended responses, human evaluators compare the original and modified question-answer pairs and choose which answer is better or whether both are acceptable.The evaluation process also includes VLM question modification, human verification, and model responses to both versions.Source: §3.1 · p. 4, Figure 2 · p. 4

4 VLM Performance under Gricean Violations

The section tests how VLMs respond to answer-preserving and hallucinated question modifiers, and compares their reasoning with human responses. Modifiers generally reduce accuracy, while models differ in how they handle unreliable information and align with humans.

4 VLM Performance under Gricean Violations

  • The study tests four VLMs on 995 VQA v2.0 questions, comparing responses to originals and modified questions across binary, quantitative, and open-ended types.Three AMT workers supplied open-ended ground-truth answers.Source: §4 · pp. 4–5

4.1 Evaluating Response Accuracy

  • Almost all modification-model and response-model combinations show decreased accuracy, with GPT answering its own modifications as the notable exception; Gemini is the most damaging modifier-model and Llava the least accurate responder.Claude has the smallest accuracy change, 4.82%.Source: Table 1 · p. 5, §4.1 · p. 5
  • McNemar’s tests find significant modifier effects in most cases, but Claude’s responses do not change significantly for three of four modification-models.Claude also has the smallest accuracy standard deviation reported for its response changes.Source: §4.1 · p. 5, Table 2 · p. 6
  • For open-ended responses, Gemini’s modifications make reasoning harder across response-models, while Llava produces the worst responses; the AMT and exact-match evaluations have an overall correlation of 0.21.For the open-ended comparison, AMT users judged whether modified-question responses were worse than original-question responses.Source: §4.1 · pp. 5–6, Table 3 · p. 6

4.2 Pragmatic Reasoning under Unreliable Information

  • The hallucination test has modification-models add false or irrelevant visual or relational details without seeing the image, which response-models can use to correct the text.This targets violations of the maxims of quality and relation.Source: §4.2 · p. 6
  • For hallucinated modifiers, unchanged responses operationalize pragmatic handling; GPT remains unchanged in around 43% of cases, versus over 60% for each other model.Figure 3 averages results over four hallucinated modifications from each VLM.Source: §4.2 · pp. 6–7, Figure 3 · p. 7
  • For binary and quantitative hallucinated questions, GPT, Gemini, and Claude match human answers better than Llava; open-ended responses have higher BERTScore F1 than constrained responses.Binary VLM answers align closely with human negative answers, but match positive answers in roughly half of such cases.Source: Table 4 · p. 7, §4.2 · p. 7
  • When a binary question is partly true, VLMs tend to answer negatively, whereas humans may answer positively depending on the partially true content’s importance in the image.The paper characterizes this contrast as more logical VLM inference versus human inference that assumes uncertainty.Source: §4.2 · p. 7

5 Comparing Human-Generated and Model-Generated Violations

The section compares VLM responses to human- and AI-modified VQA questions and measures human time-on-task, finding better VLM performance on human edits but similar human effort distributions.

5.1 Comparison of Accuracy

  • For all four models, binary-question accuracy was lower on VLM-modified than human-modified questions; Claude performed best and Llava worst on both.For open-ended questions, Llava responses to AI-modified questions were significantly worse than its responses to human-modified questions; Claude’s were judged worse in 26% of cases.Source: Figure 4 · p. 8, §5.1 · p. 8, Figure 5 · p. 8

5 Comparing Human-Generated and Model-Generated Violations

  • The comparison uses 500 human-modified binary questions and 130 open-ended inference questions sampled from a dataset of human-modified VQA questions.The open-ended sample included who, which, where, and why questions.Source: §5 · p. 8

5.1 Comparison of Accuracy

  • Human modifications had CLIPScore 0.612, closest to Claude’s 0.639 for binary questions and Gemini’s 0.644 for open-ended questions; Llava’s longer modifiers were less human-like.Llava’s modifiers averaged 3.8 words, compared with 2.1 words for human modifications; the authors summarize that VLMs handle more realistic violations more effectively.Source: Table 6 · p. 9, §5.1 · p. 9

5.2 VLM Reasoning in Relation to Human Cognitive Effort

  • Human time-on-task distributions were similar for human- and VLM-modified questions; means were 73.4 and 74.2 seconds, respectively, despite the authors’ lower-effort conclusion for VLM edits.Workers answered a modified question and simplified it by removing information they did not need; times above 500 seconds were filtered out.Source: Figure 6 · p. 9, §5.2 · p. 9

6 Conclusion

The conclusion reports reduced VLM accuracy under Gricean violations, differences from human pragmatic reasoning, and better performance on human-induced than VLM-generated modifications.

6 Conclusion

  • VLMs were generally less accurate on VQA questions containing modifiers that introduced Gricean violations.The violations could make questions ambiguous, overly detailed, or include false or irrelevant information.Source: §6 · p. 10
  • VLMs acted more logically than humans in the comparison, and performed better on human-induced than on unnatural VLM-generated modifications.Source: §6 · p. 10
  • The authors identify accounting for uncertainty in VLM-generated modifications as future work.Source: §6 · p. 10

7 Limitations

The authors identify assumptions about source questions, possible self-modification bias, open-ended evaluation bias, and unhandled response uncertainty as limitations.

7 Limitations

  • The experiments assume original human-written questions generally follow Grice’s maxims, although some may violate them even when asked as questions.Source: §7 · p. 10
  • Self-modification results may be biased if a VLM’s modification and answering prompts are not fully independent, despite using separate prompt sessions.Source: §7 · p. 10
  • Because open-ended questions lack unique ground-truth answers, the evaluation may introduce human-induced bias relative to questions with unique answers.Source: §7 · p. 10
  • VLM responses are non-deterministic, and the study does not handle this uncertainty, which could bias some results.Source: §7 · p. 10

Appendix A: LLM Prompts, Datasets, AMT Designs, Examples

The appendix prompts constrain VLM answer format and question modification, specifying response rules, modifier types, and examples across the answer-generation and modification prompts.

7.1 Prompt for answer generation from LLM

  • The answer-generation prompt asks for a mostly one-word answer based on the image, with number answers in word form and occasional two- or three-word responses allowed.Source: §7.1 · p. 11

7.2 Prompt for modified question generation using VLM

  • The question-modification prompt requests one added attribute while preserving the original question, and requires a specified output format identifying the modified question and attribute type.Source: §7.2 · pp. 11–12
  • The VLM modification rules preserve question type and target noun and allow exactly one visual or relational attribute, but not both.Examples distinguish a visual adjective before the noun from a positional phrase after it.Source: §7.2 · pp. 11–12
  • The added attribute must not change the expected answer or hint at the original answer.Source: §7.2 · p. 12

7.3 Prompt for modified generation from Llava

  • The Llava prompt casts the model as a question modifier, not an answerer, and directs it to add exactly one attribute without describing the image.Source: §7.3 · p. 12
  • The Llava prompt describes visual attributes as adjectives before the main noun and relational attributes as position phrases after it.Source: §7.3 · p. 12
  • The Llava examples show visual, relational, and adjective-based edits, including changing “Where is the dog?” to “Where is the small dog?”Source: §7.3 · p. 12

Instruction: –

The appendix specifies how modified questions and datasets were prepared, and illustrates the AMT tasks and human–VLM response comparisons. It also shows the time-on-task task design used to measure cognitive load.

Instruction: –

  • Question modifiers may add visual or relational properties, and each generated output must be one question containing one or more modifiers.The prompt requests only the modified question, without explanation.Source: Instruction: – · p. 12
  • The prompt example changes “thin” to “obese” in a question about a girl wearing a black top.Source: Instruction: – · p. 12

7.5 Datasets

  • Experiments use VQA2.0 test-partition samples to generate questions with each of four VLMs, using the image and original question, or the question alone for hallucinated modifications.Source: §7.5 · p. 13
  • Tables 7–8 count open-ended question types for VLM accuracy, while Tables 9–10 describe question types used to compare human and AI modifications.Table 9 covers binary- and quantitative-response types; Table 10 covers open-ended-response types.Source: Table 7 · p. 13, Table 8 · p. 13, Table 9 · p. 13

7.6 Question verification using AMT

  • The AMT verification task asks users to judge whether a modification is correct and correct it when it is not.Figures 7–8 show the task design; the correction is entered in a textbox.Source: §7.6 · p. 13, Figure 7 · p. 15, Figure 8 · p. 16
  • Separate AMT tasks compare open-ended answers to original and modified questions and measure time-on-task while users process violations.Figure 9 shows the answer comparison, and Figure 10 shows the time-on-task task design.Source: §7.6 · p. 13, Figure 9 · p. 17, Figure 10 · p. 18

7.7 Examples

Related papers

Loading 2610.02878v1…