PaperlayerView exact revision on arXiv

Source-linked AI summary

PVFH-Bench: Benchmarking Premise-Induced Visual Fixation Hallucinations in Vision-Language Models

Xinhao Wang, Junbo Cui, Zhenghao Liu, Yuan Yao, Xiaohua Li, Yu Gu, Ge Yu

arXiv:2610.14635v1cs.CLcs.CV

TL;DR

Object-level hallucination benchmarks do not adequately test whether LVLM reasoning is grounded in accurate visual understanding. PVFH-Bench constructs visually confusable absent targets and diagnostic subsets to study premise-induced fixation. Across evaluated models, real-object recognition can coexist with vulnerability to false premises, and explicit reasoning does not necessarily improve faithfulness.

  • Problem

    Object-level faithfulness benchmarks do not adequately assess whether reasoning-capable LVLMs ground their reasoning in accurate, reliable visual object understanding.

    Source: §1 · p. 1

  • Method

    PVFH-Bench automatically constructs visually confusable absent targets and pairs a core false-premise setting with diagnostic subsets to trace fixation-related failure.

    Source: §1 · p. 2, §3.1 · pp. 3–4

  • Results

    Models can recognize real objects yet remain vulnerable when false premises guide fixation toward visually confusable absent objects; explicit reasoning does not necessarily improve faithfulness.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    Reliable object recognition alone does not ensure faithful reasoning when a false premise directs attention toward a plausible absent object.

    Source: §6 · p. 9

    The relation-style false-premise queries do not cover all open-ended misleading premises in real interactions, and free-form responses are graded by an LLM.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large Vision-Language Models (LVLMs) have achieved remarkable progress across a wide range of multimodal tasks. However, object hallucination remains a persistent challenge, undermining model reliability in applications such as visual question answering, embodied agents, and human-AI decision support systems. In this paper, inspired by cognitive science studies on perceptual bias, we investigate a new failure mode termed Premise-Induced Visual Fixation Hallucination (PVFH). Specifically, PVFH refers to a premise-induced reasoning error in which models become anchored to visually confusable evidence, accept a false textual premise, and subsequently generate responses about an absent yet visually plausible object. To systematically study this phenomenon, we introduce PVFH-Bench, a benchmark that automatically constructs visually confusable but absent targets and pairs the core hallucination setting with diagnostic subsets to trace how models move from object recognition to Premise-Induced Visual Fixation Hallucination. Experimental results on 10 state-of-the-art LVLMs demonstrate that all evaluated models are highly vulnerable to Premise-Induced Visual Fixation Hallucination. Moreover, we find that explicit chain-of-thought reasoning can further exacerbate this failure mode by encouraging models to overinterpret weak visual cues as supporting evidence for hallucinated objects.

1 Introduction

Object-hallucination benchmarks have not adequately tested whether LVLM reasoning is grounded in accurate visual understanding. The paper defines Premise-Induced Visual Fixation Hallucination and introduces PVFH-Bench to trace this failure, finding that explicit reasoning can reinforce it.

1 Introduction

  • Existing object-level faithfulness benchmarks do not adequately assess whether reasoning-capable LVLMs ground reasoning in accurate, reliable visual object understanding.The paper connects this gap to trustworthy decision-making.Source: §1 · p. 1
  • Premise-Induced Visual Fixation Hallucination occurs when a false-premise query anchors a model to a visually confusable region and elicits an answer about an absent object.The motivating example describes a newspaper-like region treated as evidence for a nonexistent magazine.Source: §1 · p. 2
  • PVFH-Bench pairs a core hallucination setting with diagnostic subsets to trace movement from ordinary object verification to premise-induced fixation.Its targets are automatically constructed to be visually plausible but absent.Source: §1 · p. 2
  • Models can recognize real objects yet fail when false premises steer fixation toward visually confusable absent objects.The introduction reports that explicit reasoning can further reinforce this failure.Source: §1 · p. 2

2 Related Work

Prior work evaluates object hallucination and increasingly challenging false-premise scenarios, but the interaction between reasoning and fixation on plausible absent objects remains unexamined.

2 Related Work

  • VLM development has progressed from lightweight vision-language alignment to higher-resolution encoders and systems with intermediate reasoning processes.The cited examples include frozen vision encoders with LLM decoders and newer systems with explicit or implicit thinking.Source: §2 · p. 2
  • Prior hallucination benchmarks assess object mentions, existence polling, or broader multimodal performance, while reasoning under false-premise queries remains insufficiently explored.The cited examples include CHAIR, POPE, and MME for evaluation, and HaloQuest for visually challenging and false-premise-driven cases.Source: §2 · p. 2
  • Existing work does not explicitly study how a visually plausible absent object introduced by a false-premise query becomes a referential anchor during reasoning.The paper identifies this as Premise-Induced Visual Fixation, which can increase the likelihood of hallucinated responses.Source: §2 · pp. 2–3

3 Benchmarking Premise-Induced Visual Fixation Hallucination

PVFH-Bench isolates premise-induced fixation from ordinary visual recognition by pairing visually confusable absent targets with diagnostic queries and a controlled evaluation procedure.

3 Benchmarking Premise-Induced Visual Fixation Hallucination

  • PVFH-Bench is designed to separate premise-induced fixation hallucination from ordinary visual-recognition errors using plausible absent objects, true objects, and two query types.The queries test basic visual understanding and false-premise evaluation under premise pressure.Source: §3 · p. 3

3.1 Dataset Construction

  • The benchmark selects a visible, unambiguous true object and an absent, mutually exclusive, visually similar false object, retaining the most confusable qualifying pair.A visual reasoning model proposes candidate pairs; an example pairs a newspaper with an absent magazine.Source: §3.1 · p. 3
  • Eight spatial-relation templates create false-premise queries, while true-object, false-object, relational false-premise, and random-false-object control settings test distinct failure stages.The shared templates and object pairs help distinguish visual confusability from premise pressure.Source: §3.1 · pp. 3–4

3.2 Dataset Source

  • The benchmark uses 258 MSCOCO 2014 validation images, with one validated object pair per image and 258 instances in each of four subsets.Keeping the same image and pair across subsets supports controlled comparisons of recognition, fixation, premise acceptance, and the combined failure.Source: §3.2 · p. 4

3.3 Evaluation Methodology of PVFH-Bench

  • An LLM judge maps open-ended responses to correctness labels, checking semantic agreement for yes/no tasks and premise rejection for absent-object queries.Qwen3.5-Flash is used as the judge model.Source: §3.3 · p. 4

4 Experimental Methodology

The experiments assess accuracy across PVFH-Bench’s core setting and diagnostic subsets, using non-reasoning evaluations broadly and reasoning comparisons on a smaller model set.

4 Experimental Methodology

  • PVFH-Bench evaluation covers the core PVFH setting and diagnostic subsets for basic recognition, visual fixation, and false-premise pressure.Vanilla tests actual-object existence, Visual Fixation tests visually confusable absent objects, and False Premise tests relational queries about absent objects.Source: §4 · p. 4
  • All reported evaluation scores are accuracies in percentage points.The metric follows POPE.Source: §4 · p. 4
  • The evaluation spans 10 LVLMs from four model families and both open- and closed-source systems, with reasoning-mode comparisons on five models.All models are first evaluated in default non-reasoning mode to establish the main benchmark results.Source: §4 · p. 5
  • Generations use temperature 0.1, while Seed reasoning-mode evaluations use high reasoning effort.The stated setup reduces sampling variability and provides a sufficient reasoning budget for testing faithfulness under premise pressure.Source: §4 · p. 5

5 Experiment Results

LVLM accuracy falls sharply in the premise-induced visual-fixation setting, and explicit reasoning further reduces robustness. Diagnostic analyses show how false premises can redirect reasoning toward visually plausible substitutes, while greater reasoning depth does not reliably restore faithfulness.

5.1 Overall Performance

  • 43.51% average accuracy in Premise-Induced Visual Fixation Query, versus 97.36% on Vanilla, marks the sharpest degradation; visually confusable absent targets lower accuracy by 29.25 points.With templates and grading fixed, the paper reports a similar 27.99-point drop under an independent judge.Source: §5.1 · pp. 5–6
  • Qwen3.5-Plus and Seed-2.0-Lite are most robust overall, averaging 86.82% and 82.85%, while core-setting scores range from 74.81% for Qwen3.5-Plus to 1.94% for Gemma-3-4B-IT.The diagnostic columns show Qwen and LLaMA models often retain stronger accuracy on false-premise diagnostics alone, while Gemma models are already brittle there.Source: §5.1 · p. 6
  • Across five models with reasoning-mode results, average Premise-Induced Visual Fixation accuracy falls from 60.93% without reasoning to 40.16% with explicit reasoning.Qwen3.5-Plus drops from 74.81% to 36.82%, while Seed-2.0-Mini drops from 44.96% to 27.13%.Source: §5.1 · p. 6

5.2 Evaluating the Impact of False Premises on Visual Reasoning

  • False-premise queries elicit longer reasoning, but accuracy reaches its lowest point at Premise-Induced Visual Fixation along both diagnostic paths.More than two thirds of these errors occur on samples answered correctly under Visual Fixation Query; the dominant failure treats a real visible object as evidence for the absent queried object, and reasoning-chain misalignment tracks the accuracy gap.Source: Figure 3 · p. 6, Figure 4 · p. 6, §5.2 · p. 7

5.3 Analyzing the Effect of Reasoning Depth

  • Across Seed models, neither shows consistent accuracy improvement across settings as reasoning levels rise, and Premise-Induced Visual Fixation remains the most fragile setting.Stronger reasoning can give models more opportunities to follow textual traps, search for visual substitutes, and rationalize a premise-consistent answer.Source: §5.3 · p. 7, Figure 6 · p. 7

5.4 Case Study

  • In two Qwen3.5 Plus cases, reasoning turns ambiguous or absent objects into concrete spatial answers by mapping visible substitutes onto the queried category.One trace maps a newspaper-like region to a presumed magazine; another recognizes that no license plate is present but continues searching for substitutes instead of stopping.Source: §5.4 · p. 8, Figure 7 · p. 8

6 Conclusion

PVFH-Bench diagnoses premise-induced visual fixation by pairing visually confusable absent targets with diagnostic subsets. The experiments show that models can recognize real objects yet hallucinate when false premises guide fixation, and that explicit reasoning does not necessarily improve faithfulness.

6 Conclusion

  • PVFH-Bench pairs visually confusable but absent targets with diagnostic subsets to diagnose Premise-Induced Visual Fixation Hallucination in state-of-the-art LVLMs.The design shows how models can recognize real objects yet fail when false premises guide fixation toward false objects.Source: §6 · pp. 8–9
  • Reasoning often becomes longer under false premises, and traces show models turning real visual evidence into support for hallucinated targets.Source: §6 · p. 9
  • Explicit reasoning does not necessarily improve faithfulness in the false-premise setting.Source: §6 · p. 9

Limitations

The work diagnoses PVFH rather than mitigating it, and its benchmark and data impose scope boundaries. The evaluation uses controlled relation-style false-premise queries and COCO images.

Limitations

  • The study diagnoses PVFH rather than testing mitigation methods.It identifies training- or inference-time premise verification as a future research direction.Source: Limitations · p. 9
  • PVFH-Bench uses visually confusable objects and relation-style false-premise queries, so it does not cover all open-ended misleading premises in real interactions.Source: Limitations · p. 9
  • The benchmark uses COCO images; extending it to robotics, medical imaging, and remote sensing may reveal domain-specific PVFH forms.Source: Limitations · p. 9

Ethical Considerations

The benchmark uses an established public image dataset and does not collect new images or personal information. Its constrained object-similarity generation process is intended to reduce added ethical and safety risks.

Ethical Considerations

  • PVFH-Bench uses images sampled from the MSCOCO validation set and does not collect new images or personal information.Source: Ethical Considerations · p. 9
  • The generation process identifies objects already in each image and proposes visually similar but absent categories, limiting the addition of new visual content.Source: Ethical Considerations · p. 9
  • The established dataset, constrained generation, and platform safety mechanisms help reduce ethical and safety risks in benchmark construction.Source: Ethical Considerations · p. 9

A.1 Human Validation of Benchmark Construction and Grading

A stratified manual audit checked benchmark annotations and grading across four tasks. The audited annotations were mostly valid, and automatic grading showed high agreement with blind human judgments.

A.1 Human Validation of Benchmark Construction and Grading

  • The audit sampled 200 instances, with 50 from each of the four tasks, and deduplicated repeated checks of image-level annotations.Source: §A.1 · p. 11
  • Auditors checked object presence or absence and independently judged model responses before comparing human decisions with Qwen3.5-Flash labels.Source: §A.1 · p. 11
  • 95.3% of audited true-object annotations and 99.1% of false-object annotations were valid; human–Qwen agreement reached 95.5% overall.Agreement was at least 92.0% in every task.Source: §A.1 · p. 11

A.2 Re-grading with an Independent Non-Qwen Judge

The appendix tests whether key findings persist under an independent judge and documents the benchmark’s diagnostic design, evaluation procedures, and scope.

A.3 Controlled False-Premise Contrast

  • Replacing an independently generated absent object with a visually confusable one drops average accuracy from 72.76% to 43.51%, while independent grading reproduces a 27.99-point drop.Under the independent judge, accuracy falls from 70.43% to 42.44%.Source: §A.3 · p. 11

A.5 Construction and Evaluation Prompts

  • Construction prompts identify a true object and visually confusable absent counterpart, then instantiate existence and false-premise spatial queries; diagnostic prompts compare reasoning traces and classify failure sources.The failure analysis distinguishes erroneous visual evidence from premise-only assumption or other factors.Source: §A.5 · pp. 11–12

A.2 Re-grading with an Independent Non-Qwen Judge

  • Qwen3.5-Plus remains first under GPT-5.4-mini grading at 86.05 overall, changing by −0.78 points; this stable ranking argues its lead is not a Qwen-grading artifact.Seed-2.0-Lite has the largest judge-specific shift, decreasing by 4.55 points.Source: §A.2 · p. 11, Table 3 · p. 12

A.6 LLM-as-a-Judge Prompts

  • For open-ended VQA, the LLM judge checks semantic agreement on existence questions and accepts false-premise spatial responses only when they reject the absent object or refuse the forced choice.A concrete spatial relation under the false premise is marked incorrect.Source: §A.6 · p. 12

A.7 Dataset Statistics; A.8 Benchmark Scale in Context

  • PVFH-Bench contains 258 image-level object pairs and 1,032 instances across four query subsets, spanning 293 object categories; its 258-image scale supports controlled comparisons, not broad domain coverage.The categories include sports equipment, food, animals, vehicles, furniture, and household objects.Source: §A.7 · pp. 12–13, §A.8 · p. 13

A.4 Attention Evidence for Visual Fixation

  • Mean attention per reasoning token retains 93.1% of the Visual Fixation level, while answer-stage attention per token rises 17.4% under Core PVFH.The analysis measures attention to grounded true-object regions in 50 image-matched pairs.Source: §A.4 · p. 11, Table 4 · p. 13

A.10 Model Licenses; A.11 Dataset License; A.12 Annotator Information and Ethical Considerations

  • Evaluated models were used under their respective licenses; MSCOCO images and generated questions and annotations are covered by CC BY 4.0, and one volunteer annotator gave informed consent.The annotator participated in data labeling as part of an academic collaboration.Source: §A.10 · p. 13, §A.11 · p. 13, §A.12 · p. 13

A.9 POPE Comparison for Qwen3-VL

  • Qwen3-VL-8B accuracy falls from 88.84 on POPE to 58.14 on the Premise-Induced Visual Fixation Query; Qwen3-VL-32B falls from 89.11 to 58.14.The comparison covers the two Qwen3-VL variants used in the experiments.Source: §A.9 · p. 13, Table 8 · p. 17

Related papers

Loading 2610.14635v1…