CaughtCheating: Is Your MLLM a Good Cheating Detective? Exploring the Boundary of Visual Perception and Reasoning

CaughtCheating: Is Your MLLM a Good Cheating Detective? Exploring the Boundary of Visual Perception and Reasoning

This is the repo for the CaughtCheating project, in which we explore the current boundary of MLLMs and construct a hard benchmark for visual perception and reasoning.

The repo contains:

• The data for CaughtCheating Benchmark.
• The code for CaughtCheating Evaluation.

(Feel free to email Ming (Email) for any questions or feedback.)

News

• [2025/06] We released our project.

Contents

• Overview
• Highlights
• Benchmark
• Evaluation
• Citation

Overview

Recent agentic Multi-Modal Large Language Models (MLLMs) such as GPT-o3 have achieved near-ceiling scores on various existing benchmarks, motivating a demand for more challenging test tasks. These MLLMs have been reported to excel in a few expert-level tasks for humans, e.g., GeoGuesser, reflecting their potential as a detective who can notice minuscule cues in an image and weave them into coherent, situational explanations, leading to a reliable answer. But can they match the performance of excellent human detectives? To answer this question, we investigate some hard scenarios where GPT-o3 can still handle, and find a common scenario where o3's performance drops to nearly zero, which we name CaughtCheating. It is inspired by the social media requests that ask others to detect suspicious clues from photos shared by the poster's partner. We conduct extensive experiments and analysis to understand why existing MLLMs lack sufficient capability to solve this kind of task. CaughtCheating provides a class of challenging visual perception and reasoning tasks with great value and practical usage. Success in these tasks paves the way for MLLMs to acquire detective-level visual perception and reasoning capabilities.

Highlights

• We systematically evaluate the limits of current MLLMs in visual perception and reasoning, analyzing how they solve various complex tasks via sophisticated reasoning strategies, and identify a common scenario where even advanced models like o3’s performance drops to nearly zero.
• We present CaughtCheating, the first benchmark specifically designed to assess the ability to actively search and detect subtle, context-dependent suspicious clues in real-world images. Most human annotators and state-of-the-art agentic MLLMs struggle to succeed on CaughtCheating tasks, highlighting the lack of detective-level exploration skills.
• We analyze why even the most advanced agentic MLLMs fail on CaughtCheating. Inspired by the Guided Search theory, we find that these models often lack awareness of what to search for and how to relate observed details to the query. Our findings offer insights into both the construction of more challenging benchmarks and the limitations of existing MLLMs.

Benchmark

Exploring the Boundary of Visual Perception and Reasoning

Demonstration of GPT-o3’s multimodal visual-reasoning breadth. (a) Visual search: locating Waldo in a densely populated illustration. (b) Visual search for camouflage: spotting a nearly invisible copperhead snake hidden among dry leaves. (c) GeoGuessr: identifying the upper terminal of Chair 1 at New Mexico, and estimating its latitude/longitude from a single image. (d) TimeGuesser: dating the photograph by matching architectural signage and period vehicles to museum and heritage records. These examples highlight o3’s strong visual perception and reasoning capacity across various visual tasks that most humans can not accomplish.

Annotation Example

An example of the annotation for the "Clued" category. Each image is annotated with a general question assessing overall suspicion and decomposed questions focused on a deterministic clue (here, the feminine bow hair accessory). Decomposed questions include perception-based inquiries (clue identification) and reasoningbased inquiries (social implications and contradictions), all annotated with the expected answer "yes".

Performance

The accuracies, IoU on the Clued category, and the accuracy on the Unclued category, and the overall precision, recall, and F1 score. Models are grouped by parameter size and type (open-source vs. proprietary). Clued Acc and IoU represent the capability of MLLMs to identify the suspicious clues, which directly reflects the MLLMs’ visual perception and reasoning abilities. Even the best performing model, GPT-o3, only achieves 26.0% accuracy and 17.2% IoU, indicating the current boundary of MLLMs’ capabilities. Unclued Acc represents the capability of MLLMs to not generate any suspicious clues if the image is unclued. F1 score shows the overall capability of MLLMs on CaughtCheating, where GPT-o3, achieves only 23.9%. The highest F1 score is 23.9%, which is much lower than the human performance, indicating the current boundary of MLLMs’ capabilities.

Model	Clued		Unclued	Overall
	Acc ↑	IoU ↑	Acc ↑	Precision ↑	Recall ↑	F1 ↑
LLaVA-v1.6-Mistral-7B	0.0	0.0	82.0	0.0	0.0	0.0
LLaVA-OV-7B	2.0	0.0	52.0	4.0	2.0	2.7
Qwen2.5-VL-7B	2.0	3.9	66.0	5.6	2.0	2.9
InternVL2-8B	0.0	0.0	76.0	0.0	0.0	0.0
InternVL2.5-8B	0.0	0.0	72.0	0.0	0.0	0.0
LLaVA-1.6-Vicuna-13B	0.0	0.0	72.0	0.0	0.0	0.0
InternVL2-26B	2.0	1.8	10.0	2.2	2.0	2.1
InternVL2.5-26B	0.0	0.0	80.0	0.0	0.0	0.0
InternVL2.5-38B	2.0	0.0	76.0	7.7	2.0	3.2
InternVL2-40B	4.0	0.7	12.0	4.4	4.0	4.2
InternVL2-72B	4.0	0.8	16.0	4.5	4.0	4.3
InternVL2.5-72B	2.0	0.8	80.0	9.1	2.0	3.3
LLaVA-OV-72B	0.0	1.3	72.0	0.0	0.0	0.0
GPT-4o	4.0	1.0	54.0	8.0	4.0	5.3
Gemini-2-flash	10.0	0.0	6.0	9.6	10.0	9.8
Gemini-2.5-flash	18.0	5.1	22.0	18.8	18.0	18.4
Gemini-2.5-pro	20.0	15.1	22.0	20.4	20.0	20.2
GPT-o3	26.0	17.2	8.0	22.0	26.0	23.9
Human	56.0	/	68.0	63.6	56.0	59.6

Case Study

Case studies of the models’ performance on the CaughtCheating examples. 3 representative models are selected, including GPT-o3, GPT-4o and InternVL2.5-1B, and 3 images are selected: (a) A difficult Clued image, (b) An easy Clued image, and (c) An Unclued image. The models’ responses are truncated for better visualization.

Evaluation

Folders

The dataset is structured in a flat, repeatable format for easy evaluation. The folder layout is:

dataset/
├── true_images/
│   ├── true_1.png
│   ├── true_2.png
│   └── ...
├── false_images/
│   ├── false_1.png
│   ├── false_2.png
│   └── ...
└── data_info.json

Key points:

• true_images and false_images contain images labeled as true or false respectively.
• data_info.json provides metadata, instructions, labels, and associated URLs for each image.

data_info.json Format (Per-Sample)

{
    "data_path": "path/to/image.png",   // Relative path to the image file
    "label": true,                      // True or false indicating ground truth
    "scene": "scene description",       // Scene context (e.g., hotel, dining)
    "target_partner": "partner type",   // Targeted partner (e.g., boyfriend, girlfriend)
    "main_ins": "Main instruction/question", // Primary instruction or query
    "detm_cue": "Determinate cue text",       // Determinate cue clearly indicating suspicion
    "non_detm_cue": [                   // List of non-determinate cues (optional)
        "cue 1",
        "cue 2"
    ],
    "p_sub_ins": [                      // Perception sub-questions
        "sub-question 1",
        "sub-question 2"
    ],
    "p_sub_ans": [                      // Answers to perception sub-questions
        "Yes",
        "No"
    ],
    "r_sub_ins": [                      // Reasoning sub-questions
        "sub-question 1",
        "sub-question 2"
    ],
    "r_sub_ans": [                      // Answers to reasoning sub-questions
        "Yes",
        "No"
    ]
}

Quick Start

Please follow these steps to prepare your dataset, run inference, and evaluate results:

Step 1: Convert Annotations from JSON to JSONL

Convert annotations from standard JSON format (data_info.json) to JSONL format.

python code/json2jsonl.py --input_file dataset/data_info.json --output_file dataset/data_info.jsonl

Key arguments:

• --input_file: Path to your original JSON annotations (data_info.json).
• --output_file: Path to save converted JSONL annotations (data_info.jsonl).

If your annotations are already in JSONL format, you can skip this step.

Step 2: Run Model Inference

Perform inference using your selected model (e.g., GPT-4o). The inference script processes input annotations and generates predictions.

python code/inference.py --root_dir dataset --save_dir gpt4o_results.json --datatype json --model_name gpt-4o --api_key YOUR_API_KEY

Key arguments:

• --root_dir: Directory containing data_info.jsonl and image files (default: ./dataset).
• --save_dir: Directory to save inference results (default: gpt4o_results).
• --datatype: Indicate how images are referenced (json by default).
• --model_name: Specify the OpenAI model name to use for inference (e.g., gpt-4o, gpt-4o-mini).
• --api_key: Your OpenAI API key (required).

Step 3: Evaluate Inference Results

Evaluate inference outputs against ground truth to compute metrics and detailed per-sample results.

python code/eval.py --input_file gpt4o_results.json --gpt_model gpt-4.1 --output_file evaluation_gpt4o_results.json --api_key YOUR_API_KEY

Key arguments:

• --input_file: Path to your inference results (inference_results.json).
• --gpt_model: Identifier for the model used (gpt-4.1).
• --output_file: Path to save evaluation metrics and detailed results (evaluation_results.json).
• --api_key: Your OpenAI API key for accessing evaluation services.

Ensure your OpenAI API key or necessary authentication tokens are properly configured before running these commands.

Ensure the API key or necessary authentication tokens are correctly configured before running the evaluation.

Citation

Please consider citing our papers if you think our code or data are useful. Thank you!

@article{li2025caughtcheating,
  title={CaughtCheating: Is Your MLLM a Good Cheating Detective? Exploring the Boundary of Visual Perception and Reasoning},
  author={Li, Ming and Wang, Chenguang and Liang, Yijun and Wang, Xiyao and Zhou, Yuhang and Wu, Xiyang and Zhang, Yuqing and Zhang, Ruiyi and Zhou, Tianyi},
  journal={arXiv preprint arXiv:2507.00045},
  year={2025}
}