Research map / Vision-language and multimodal / Multimodal LLMs

LOKI: A Comprehensive Synthetic Data Detection Benchmark using Large Multimodal Models

ICLR 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Unified any-modality language models · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces LOKI, a multimodal benchmark (video, image, 3D, text, audio; ~18K questions across 26 subcategories) that evaluates LMMs and LLMs on synthetic data detection via judgment, multiple-choice, abnormal detail selection, and abnormal explanation tasks.

Problem

Existing synthetic data detection benchmarks focus on binary authenticity classification for traditional detectors or a single modality, and do not evaluate LMMs' ability to detect and explain synthetic data across modalities.

Approach

A cross-modal benchmark with fine-grained human anomaly annotations and paired real/synthetic data that tests LMMs on judgment, multiple-choice, anomaly selection and open-ended anomaly explanation, plus a bias index (NBI) and difficulty levels from human performance.

Key results

• GPT-4o achieves 63.9% overall judgment accuracy and 73.7% multiple-choice accuracy (Claude-3.5-Sonnet 74.8% on text-including multiple choice), and 75.3% on abnormal detail selection and 72.6 on abnormal explanation.
• Humans average 76.0% on judgment and 86.4% on multiple choice; the best LMMs trail human performance by roughly 10 points.
• Open-source audio LMMs perform near random (around 49-51% judgment accuracy), and GPT-4o accuracy drops to 44.4% on hard questions.

Measured effects reported in the paper

Surprising findings

• GPT-4o accuracy on hard questions (44.4%) falls below random selection level.
• Chain-of-thought prompting sharply degrades LLaVA-OV-7B on image and 3D judgment (56.6 to 18.8), while few-shot prompting does not help most models.

Limitations

• Dataset access requires a download request to authors, to prevent misuse.
• Abnormal detail selection and explanation rely on GPT-4o-generated questions and GPT-4 scoring, with manual verification.
• Expert detectors were evaluated only on a few modalities; 3D was not evaluated for expert models.

Experimental setup

Datasets: OmniObject3D GPTEval3D ASVSpoof2019 CtrSVDD DCASE2023 Task 7 MusicCaps M6Doc M3DSynth GenImage Mixset CVUSA VIGOR

Benchmarks: LOKI

Metrics: accuracy Normalized Bias Index GPT-Score

Baselines: Human (medium) AIGVDet AIDE RADAR-Vicuna-7B AASIST Random Choice

Builds on: FakeBench VANE AIGVDet AIDE

Closest papers by meaning

• Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation (NeurIPS 2025 · similarity 0.89)
• Learning from Reasoning Failures via Synthetic Data Generation (AAAI 2026 · similarity 0.87)
• Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.85)
• SNIFFER: Multimodal Large Language Model for Explainable Out-of-Context Misinformation Detection (CVPR 2024 · similarity 0.85)
• CapsFusion: Rethinking Image-Text Data at Scale (CVPR 2024 · similarity 0.85)
• When Understanding Becomes a Risk: Authenticity and Safety Risks in the Emerging Image Generation Paradigm (CVPR 2026 · similarity 0.85)