Source-linked AI summary
Beyond What Meets the Eye: Unveiling Situational Illusions for Multimodal Large Language Models
Zhiming Yang, Zhuoxi Xiong, Donglin Zhou, Wenjun Wei, Shiyao Cui, Jinqiao Shi
TL;DR
MLLMs must infer physical states when real-world appearances are misleading, but evidence on this situational-illusion problem is limited. The paper develops a taxonomy and MSIBench, evaluates 27 model configurations, and introduces prompting and SFT mitigations. Models remain vulnerable, while the proposed methods improve performance by up to 20%.
Problem
Situational illusions make naturally context-induced appearances misrepresent physical states, a problem insufficiently explored in MLLM contextual understanding and reasoning.
Method
The paper builds a where-what-how taxonomy and MSIBench, then uses prompting for closed-source models and SFT for open-source models.
Results
27 model configurations show substantial vulnerabilities, with below 63% overall average accuracy and improvements of up to 20% from the mitigation methods.
Takeaways & Limitations
Systematically inspecting and reasoning over visual evidence offers a practical path toward more reliable multimodal decisions in real-world situations.
Takeaways & Limitations
The study identifies observation and grounding failures as constraints on reliable interpretation, with grounding failures accounting for 33.68% of wrong results.
Abstract
from arXiv · showhide
Real-world situation appearances can deviate from their underlying physical states, challenging the reliability of multimodal large language models (MLLMs) in practical applications. In this paper, we term this phenomenon situational illusions and investigate: (1) how MLLMs perform under such illusions, and (2) how to mitigate the limitations. We first develop a comprehensive where-what-how taxonomy that characterizes where situational illusions occur, what targets they take, and how they arise. Building on this taxonomy, we introduce MSIBench, a benchmark designed to assess the discrimination, understanding, and reasoning capabilities of MLLMs under situational illusions. Evaluations of 27 model configurations reveal that current MLLMs are highly vulnerable to these illusions and exhibit 6 typical failure modes related to visual observation, grounding, and reasoning. To mitigate the limitations, we build on the core idea of systematically inspecting and reasoning over visual evidence for contextual understanding, developing prompting for closed-source models and supervised fine-tuning for open-source models, respectively. These two simple yet effective methods improve model performances by 20% at most, suggesting a practical path toward more reliable multimodal perception and reasoning in complex real-world environments.
1 Introduction
Situational illusions arise when naturally context-induced visual appearances misrepresent physical reality, challenging MLLMs’ contextual interpretation and decision-making. The paper introduces a taxonomy, MSIBench, evaluations, and mitigation strategies to study and address this problem.
- Motivation: Situational illusions occur when visual appearances diverge from underlying physical states, potentially causing unsafe action plans.A mug appearing upside down may prompt an MLLM to flip it before inserting a straw, spilling milk.
- Motivation: The paper identifies situational illusions as an underexplored problem distinct from deliberately designed visual illusions and misleading structured-data cues.Existing work generally assumes visual appearance matches reality or studies designed illusion patterns and adversarial charts or tables.
- Contributions: The study develops a where-what-how taxonomy to provide a framework for analyzing where illusions occur, what they misrepresent, and how they form.The taxonomy is paired with systematic investigation of MLLM performance and mitigation.
- Contributions: MSIBench evaluates illusion discrimination, understanding, and reasoning, with adversarial settings included to stress-test MLLMs.The benchmark is constructed to reveal model limitations and inform improvement.
- Findings: 27 model configurations achieve below 63% average accuracy overall and below 45% on action planning under situational illusions.These evaluations reveal substantial vulnerabilities across the benchmark.
- Mitigation: Prompting and supervised fine-tuning improve model performance by up to 20% by encouraging attention to visual evidence.Prompting targets closed-source models, while SFT targets open-source models.
2 Taxonomy
The taxonomy organizes situational illusions by where they occur, what is misperceived, and how the illusion forms. It covers everyday scenarios, target types, and mechanisms ranging from perceptual ambiguity to cognitive bias.
- Overview: The taxonomy characterizes situational illusions across scenarios, misperceived targets, and formation mechanisms.These dimensions describe where illusions occur, what is misperceived, and how they arise.
- Scenarios: where situational illusions occur: Representative scenarios include small indoor enclosures, large indoor venues, urban outdoor scenes, and natural outdoor scenes.Examples range from classrooms and offices to concert halls, streets, deserts, and beaches.
- Targets: what is misperceived: Misperceived targets comprise properties, relations, and context.Properties include size, quantity, color, shape, orientation, and material; relations include identity, pose, and adjacency.
- Targets: what is misperceived: Context targets include scene plausibility, layout, and reference frame, extending misperception beyond individual entities.These categories concern whether scenes seem artificial, resemble familiar configurations, or are distorted by scaling or viewpoint.
- Mechanisms: how illusions form: Formation mechanisms span boundary ambiguity, angle disruption, perspective misalignment, dimension confusion, and experience-driven bias.The mechanisms range from low-level perceptual ambiguity to learned expectations that distort interpretation.
3 MSIBench Construction
MSIBench is constructed from curated real-world illusion images, structured metadata, and tasks targeting discrimination, understanding, and action planning. Quality checks and paired prompting settings support systematic evaluation.
- Data collection: MSIBench construction begins with in-the-wild images gathered from online platforms and existing illusion resources.The sources include online communities and IllusionBench+.
- Data curation: Manual filtering removes unrealistic, apparatus-dependent, ambiguous, and sensitive samples before benchmark assembly.The resulting collection is intended to provide high-quality real-world illusions.
- Metadata and quality control: Each image receives metadata describing the illusion, underlying real-world state, scenario, misperceived targets, and formation mechanism.Gemini-3.1-pro-preview (Think) generates initial metadata, which is then checked and revised.
- Evaluation tasks: T/F questions test illusion discrimination, open-ended questions test situational understanding, and action planning tests multi-step reasoning about physical scenes.Action planning requires sequences of two to six actions based on object states, constraints, consequences, and correlations.
- Benchmark composition: 904 unique images yield 3,723 image-text pairs: 1,808 T/F questions, 1,808 open-ended questions, and 107 action-planning tasks.The benchmark also provides paired vanilla and adversarial settings for T/F and open-ended questions.
4 Experiments
Evaluations across 27 model configurations show that MLLMs remain vulnerable to situational illusions, with performance varying by task, descriptions, taxonomy target, and failure mode. The results motivate evidence-focused mitigation by identifying weaknesses in observation, grounding, and reasoning.
- Evaluation Setup: Evaluations use 27 model configurations with accuracy as the primary metric across T/F, open-ended, and action-planning tasks.T/F responses use label matching, while open-ended and action-planning responses are assessed for physical-scene correctness.
- Main Results: The highest overall average accuracy remains below 63%, while action-planning accuracy falls below 45%; closed-source models outperform open-source models across all tasks.The largest average closed/open-source gap approaches 16% on action planning.
- Main Results: Increased reasoning effort yields limited, task-dependent changes: GLM-4.6V-Flash improves 8.41% on action planning, whereas Gemini-3.5-Flash declines by up to 5.61%.The authors attribute inconsistent effects partly to unnecessary assumptions and distraction from the actual scene.
- Main Results: Performance generally declines across task transitions, including Qwen3.5-2B’s ∆t/f = −22.85 and Grok-4.1-Fast (Think)’s ∆open = −34.10.The paper links these gaps to increasing task difficulty and possible reliance on superficial cues rather than underlying physical-state interpretation.
- Main Results: Adversarial illusion descriptions generally reduce accuracy, with open-source models particularly susceptible compared with closed-source models.The results suggest that adversarial descriptions can bias models toward misleading cues instead of visual evidence.
- Performance Across Taxonomy: Across scenarios, the highest accuracy remains below 59% and scenario variation reaches only 7.36 points, while context is harder than property and property exceeds context by up to 13.96% on action planning.Context requires broader scene integration than localized property judgments; angle disruption is the hardest mechanism and boundary ambiguity the easiest, with action-planning gaps reaching 31.65%.
- Failure Modes: Failure modes progress from observation errors to grounding errors and scene-level reasoning errors.Observation failures include evidence mismatch, insufficiency, and overload; grounding failures misassociate correctly identified evidence; reasoning failures misinterpret spatial relations or reference frames.
5 Mitigation Strategy
The paper mitigates situational-illusion failures by structuring visual inspection and reasoning through prompting, then strengthening general capabilities through supervised fine-tuning.
- Prompting Mitigation: Guided prompting structures processing into visual observation, contextual grounding, and reasoning about the underlying physical state.The method directs models to inspect evidence, ground objects and relationships, then reason before answering.
- Prompting Mitigation: Prompting generally improves performance, but gains vary across tasks and are larger for closed-source than open-source models.Some open-source models show only marginal gains or slight degradation, indicating prompting alone may be insufficient.
- Supervised Fine-tuning Mitigation: Supervised fine-tuning targets object localization, scene reconstruction, and action planning rather than training directly on illusion tasks.The SFT dataset preserves general utility by covering spatial relations, scene description, and action-sequence reasoning.
- Supervised Fine-tuning Mitigation: All evaluated models gain over 10 average points, with Qwen3.5-9B achieving a maximum improvement exceeding 20%.Testing uses 304 remaining MSIBench images, including 107 for action evaluation, with LLM-as-a-Judge assessment and human verification.
6 Related Works
Related work covers MLLM perception, reasoning, and decision-making in real-world situations, alongside research on visual illusions and their vulnerabilities. The paper distinguishes situational illusions from deliberately designed visual illusion patterns.
- MLLMs in Real-world Situations: Research evaluates MLLMs’ fine-grained perception, everyday reasoning, cross-image evidence integration, and risk-aware decision-making in real-world environments.Studies also address embodied task planning, hazardous instructions, and hazard identification.
- Visual Illusions: Visual-illusion research examines classic perceptual phenomena and whether MLLMs exhibit related vulnerabilities, including pareidolia.These studies primarily focus on specific visual phenomena rather than naturally context-induced situational appearances.
7 Conclusion
The paper presents a taxonomy and MSIBench for systematically studying situational illusions in MLLMs, identifies vulnerabilities and failure modes, and develops two mitigation strategies.
- Conclusion: The taxonomy characterizes situational illusions, while MSIBench evaluates 27 model configurations and reveals substantial vulnerabilities and key failure modes.The benchmark and taxonomy together support systematic analysis of the phenomenon.
- Conclusion: Prompting and supervised fine-tuning significantly improve models’ ability to handle situational illusions.The paper identifies robust multimodal reasoning for embodied agents as a direction for future work.